For half a year now I’ve been writing this newsletter mostly from the machine shop’s perspective. On Wednesday I was at ProShop’s Cutting Edge in Schaumburg, and this time the buyers were on stage and shop owners were in the seats. I took a lot of notes. And, not surprisingly, not much of what they said was about price.

The thing that got me was in the Space Force session. Somebody asked why it is so hard for small shops to get into their programs, and the colonel on the stage said two things. First, Space Force has basically zero visibility into these shops, and the shops have zero visibility into Space Force.

Second, he gave an example of how they work with new VC-funded companies, and how much easier it is for those companies to take net 90 and longer terms than for an average machine shop, simply because they have the money to buy material up front and wait. I didn’t record it, so this is not word for word, but that was his example. The rest is my read, and a few friends who were there read it the same way. So basically, before anybody even looks at who does better work, the first question is who can wait for the money the longest.

This isn’t just one colonel’s example. Roger Atkins, the president of NTMA, wrote in the association’s magazine in October 2024 that small shops are “held to 30-60 day terms for materials and outside service expenses” while “the big players operate with 60, 90, 120 or 180 day payment terms,” and that one OEM executive told him they had “just changed their payment terms to Net 120 days.” Meanwhile the DoD itself pays small business primes with a goal of 15 days. So the 120 days doesn’t come from the government. It gets added somewhere between the prime and your shop.

They Can’t See You

This visibility problem is not new. Erich Hernandez-Baquero, the Air Force’s deputy assistant secretary for space acquisition, told reporters in September that “there are lower, subtier elements of that supply chain that cross cut different mission areas... because those are bottlenecks.” The colonel in Schaumburg explained how it actually works. The RFPs are written so that space parts and ground support equipment are separate, and the ground parts should not carry the full compliance load. But then the primes do what he called “lazy flow”: they push all the requirements down the chain anyway. So a mid-tier shop gets accounting, technical and manufacturing requirements that were written for a different part. IDA checked this for Congress back in 2017 and didn’t find anybody doing it on purpose. It’s just 963 clauses, and primes pass them down as is. His advice was to “respectfully push back,” and he said the OTA contracts Space Force runs through the Space Enterprise Consortium let both sides trade requirements. Good advice. But how does a 20-person shop know which requirement it is allowed to question?

And it goes the other way too. Paul Van Metre posted last month that primes are telling some shops they should charge $380 an hour for 5-axis work. 45 people, shop owners and buyers, argued about whether anybody actually pays that. Hans Keisler, who owns Motor Control Technology, said this:

There’s a big difference between what a prime says a capable 5-axis shop should charge and what procurement will actually accept when the PO hits the table.

His guess for real awarded work was $100 to $150. Eden Cisneros, who is on the buyer side, said she never said $380 and she works with $150 to $180 for a Haas UMC 500. And Jason Premo of Acclaim Aerospace told the story that explains the whole thing. Back in the 2010s Boeing “threw up” at his $175 an hour transparent quote for titanium parts, and his mentor told him to move the margin into overhead and show $150 instead. So the prime tells you to charge more, procurement pays you less, and you end up hiding the difference in overhead. After a few years of this nobody knows the real number anymore. Including the prime.

And a side note. This is, of course, the reason why the common sense pricing that online-first shops use for parts just doesn’t work in an aerospace shop. You have to factor in what you know about this one customer: how they behaved on the last job, what the program rules are, and many other things. It is still possible, technically. But all of that intel has to live at the shop level, in some system, and that’s a real lever.

Today it all sits in the head of the shop owner, who is also the business development guy, the sales rock star, the best negotiator and maybe even the lobbyist.

What Yes Looks Like

The session about getting a seat at the prime’s table was two guys who did it together. And oh boy, they delivered. The room was laughing and taking notes at the same time.

Dan Echternkamp of AeroGrowth spent about 15 years in machining, mostly selling, and now he coaches shop owners on how to get into aerospace. Ryan Nagle was his buyer, first at SpaceX before Falcon landed, then at the hardware companies he started after SpaceX. Ryan’s first rule got a laugh in the room, and then everybody went quiet: the best way to become a strategic partner is to get your buyer promoted.

When an RFQ hits your inbox... give the buyer optionality. Hey, I need pricing, 5, 50 and 200. Here’s the cycle times, break out the NREs. Make them look like a fantastic buyer to their engineering colleagues, because those guys are going to block their promotions or give them the promotions.

Also, redline the drawing before they ask you to. And when you are late, send a photo of the part. Phone camera, coolant everywhere, doesn’t matter, just put your logo on it. He said “a picture of the part with bad news feels like generally neutral news,” and that photo gets forwarded around a company that is a hundred times bigger than you.

Dan had a story about this. He once told his buyer that a CMM delivery was pushed by two weeks. The buyer made one phone call and the machine showed up the next day. “You’re a billion dollar company. We’re all 5, 10, 15.”

Ryan’s second point is basically what I wrote in the lanes issue, just from the buyer’s side. If the part is not what you are good at, don’t quote it high hoping they go away. Tell them why you are passing. A crazy high quote reads as “I just got screwed by a sales guy.” But “this runout is too tight, I don’t have a steady rest” reads as a shop that knows exactly what it does. His words:

If somebody comes to me and says, I’ll cut anything that fits in my machine, that’s the kiss of death. There’s a hundred ways to make a part. 98 of them won’t make you margin.

Dan said the same thing about websites (my favorite topic, I’ve been a web designer in the past). Aluminum, stainless, Inconel, plastics, mill, turn, Swiss, EDM, one long list, “that to me says there’s no focus, and no focus means slower speeds at higher prices.” Put one niche up front, something like “5-axis complex, 12 to 18 inch cube,” and keep everything else in the background. Ryan had one more thing that kills it for him if the shop wants new-space work: an F-35 picture on the website. He reads it as long lead times and old processes. What he likes to see is a Formula SAE team photo, because race people know that “if it’s Sunday and you’re late, you don’t get to say sorry.”

Two more things from Ryan. He audited each supplier he ever worked with, flew people out, paid for hotels, “and they’ve all delivered late.” So he stopped trusting the audit. What he looks at now is whether you teach him stuff without being asked, and whether “end of day Tuesday” means Tuesday. And you don’t need AS9100 to start. Space Force’s own supplier standard, SMC-S-003, says certification for sub-tiers is “preferred,” not required, and the prime takes responsibility for an uncertified supplier through audits and records. Ryan’s way in: open an account with a big OEM as their R&D supplier, take development work for a couple of years, let their supply chain engineers audit you and give you the majors and minors, and then switch to production. Dan talked about the same thing on MakingChips in March. When he asked the room who is AS9100 certified, most hands stayed down.

Scarce, Not Rare

At the end of the day Ryan said one thing that I think matters most. He doesn’t have a problem finding capacity. He has a problem finding the right match.

Every machine that comes out, that’s 120 more hours a week of machining capacity. So it is not rare, but it is very scarce... The junior varsity squad looks at a drawing and says, this is the shop to make it. The varsity squad says this is the exact machine to make it and this is the supplier that has it. That two-way match is what drives the scarcity.

And machines are coming in. US orders for manufacturing technology were a record $5.74 billion in 2025, up 22.5%, and contract shops, the biggest buyers, were up 19.1%. What he asked from the room: once you’ve done the NRE, the programming and the setup on a part, you are “the ultimate sourcing manager” for that part. He would much rather hear “I’m over my skis, send this run to Joe down the street” than silence.

Christian Keil from a16z wrote the same week that the hard part of hardware is the machinist picking stock, workholding and strategy, “an artisanal process,” and that “the AI labs need to acquire process knowledge” and can’t find the data. So the buyer at the prime and the buyer at the AI lab want the same thing from a shop: a clear picture of what the shop can do this week.

And then he said something shop owners don’t like to hear. A bigger customer pays for a rush, you take it, and the smaller customer’s parts slip three days and nobody emails them. “If my burn rate is 100 grand a day and you’re three days late, you may think that’s not a big deal, but I just lost 300 grand to somebody who wanted to make 25 grand.” He said he doesn’t need to be made whole. A thousand dollars off the PO and a note is enough, so he knows somebody noticed. “Anybody in this room wants to deal with that, you’re instantaneously my supplier.” I don’t think a single shop in that room does this today.

The Terms Decide

Back to the Space Force thing, because I think this is the important part. The colonel said something I hear from shop owners all the time: “The machine shops pinch. You have to pay the distributor for the bill. They don’t give you net 120 terms. You have to pay the anodize house.” Kevin Michaels did the math in Aviation Week back in 2023. A Tier 3 make-to-print shop pays for aluminum within 30 days, makes the parts, passes inspection, and only then the 120-day clock starts. Six months or more from cash out to cash in. A VC-funded supplier covers that gap with investors’ money, and a 30-person shop covers it with its own cash. So yeah, when the colonel gives VC-funded suppliers as his example, I don’t think it’s about machining quality at that point. It’s about who can wait.

And the terms can change on you without anybody telling you. A shop owner I talked to the Sunday before Schaumburg did a risk audit this year and PO entry came out as his number one operational risk. A customer changed the terms on a PO from net 30 to net 60, nobody caught it at entry, and in 30 days it took $300,000 of cash out of the business. His fix, and I would put this on the wall: let the software read the PO against the quote and flag what changed, and then a person reads the flags.

My Take

If I put all three sessions and the LinkedIn thread together, buyers asked for four things. Not surprisingly, lower price is not one of them. A niche they can repeat back to you in one sentence. A real picture of your capacity, including when you are full. Terms you can actually carry without hiding margin in overhead. And a record of what you did for them, down to the photo of the late part. On Monday I wrote that a lane is a shop that says no. The buyers basically said the same thing from their side. The shop that shows its niche, its lead time and its backlog is the shop they can see. The shop that lists every material and every process and quotes every RFQ, they just don’t see it.

The one thing I can’t fix with advice is the balance sheet. The supplier in the colonel’s example can wait 120 days, and no amount of transparency changes how much cash a 30-person shop has. That’s a financing problem. But that is how the work gets awarded in practice, whether anybody says it from the stage or not. So if you want prime work, you either find a way to carry those terms (the lanes did it with investors’ money), or you price the terms the same way Ryan says to price quantities: here is net 30, here is net 90, here is the number for each. I think, and I could be wrong here, that a buyer who is ready to take a thousand dollars off a late PO will also take a shop that prices its terms.

Watching

• Whether a prime or a new-space buyer puts the expedite rebate in writing. Paul Van Metre’s thread is where it would show up first.

• Space Force’s subtier supply chain plan. Hernandez-Baquero talked about multiyear assurances for suppliers in September. If that ever reaches a 30-person shop, the balance sheet filter gets weaker.

• MakingChips episode 23 with Dan, for the AS9100-later path in his own words.

Const

Two weeks ago Caleb Chamberlain of OSH Cut wrote the post about agents buying parts through MCP servers. Ask me a few years ago and I would never have believed AI could move this fast. Now? Well, I can see it.

The next day he wrote the post that I think matters more:

We need dozens of productized manufacturing “lanes.” Today, OSH Cut and a few other newer manufactures provide sheet metal in this productized way. But we need probably a dozen additional lanes for different types of CNC machining, tolerance requirements, mold creation, textiles, plastics, etc.

This isn’t going to be created by some new AI venture, it’ll be a bunch of factories operated by people.

So… manufacturing services in a productized way. I like how it sounds. I decided to dig deeper into it. What is a lane, who built one, who died building one, which processes still don’t have one. Here’s where I landed. A lane is a shop that says no. The owner writes down what the shop won’t make, and the website won’t let you order it. The shops that died did the opposite: they gave a fixed price on parts where the file didn’t show the real cost. And a job shop doesn’t need to become a lane. It’s going to be their customer.

This one is long. So hold on… and read.

28 Years of Saying No

The story is older than OSH Cut or SendCutSend and it started in software, not in a shop. In 1998 a company in Berlin called Schaeffer AG put its Front Panel Designer online: you draw your aluminum panel in their free program, the program prices it, you order. You can’t draw anything the machine can’t cut, because the program won’t let you. (They still run it today, from Berlin and Seattle, as Front Panel Express.)

In 2000 Ron Hollis in Atlanta launched Quickparts, and that December filed the patent on “firm binding price quotations for custom manufactured parts” from an uploaded file. Protomold’s Larry Lukis started in Minnesota in 1999, and the 2003 Machine Design story on him reads like a rules page: two-part aluminum molds only, no EDM, software that flags sharp corners and undercuts, five days, about $1,995 for 25 parts. Brad Cleveland, who ran it, said the whole business in one line:

If they can live without details such as sharp corners, we can build the part.

Then it spread one process at a time. eMachineShop in New Jersey, 2004, with its own free CAD that throws an error when you draw a feature it can’t machine. Big Blue Saw near Atlanta, 2006, waterjet from a DXF. Protomold added Firstcut in 2007: CNC, plastics only, three-axis only, parts smaller than 10 by 7 by 3 inches, one business day. Ponoko the same year, laser cutting from a vector file, three sheet sizes and nothing else. Shapeways in 2008, four materials, priced by the cubic centimeter. In 2009 a guy in Portland named Laen started OSH Park: purple mask, ENIG, sets of three, five dollars a square inch, no blind or buried vias. That was his menu.

Here’s the one a lot of people in sheet metal forget. In 2013 RAPID Sheet Metal in Nashua, New Hampshire shipped SolidQuote, a plugin that priced your part inside SolidWorks while you were still drawing it. Jay Jacobs, who ran RAPID, took 3D files only, no prints, when the shops around him were asking for drawings. Protolabs bought RAPID in 2017 for $120 million. This NO was worth nine figures.

Then 2013, the marketplace year: Xometry, Fictiv, 3D Hubs, MacroFab, Tempo and Plethora, founded within months of each other, with venture money and the idea that software could price anything and a network could make it. Then in November 2018 the bootstrapped sheet metal shops arrived: SendCutSend took its first online order on the 1st, and OSH Cut, founded that July, released its quoting app the same month. Fabworks in Ventura went online in April 2023. Tube lasers got instant prices in 2023, tube bending in 2024, castings and same-day CNC in 2025, welding late last year, raw metal this year. And Charlie Hushek, a fifth-generation heat treater at Phoenix Heat Treating, said last Tuesday he’s building the heat-treat version.

One name that is not on that chart on purpose: MFG.com. Mitch Free started it in Atlanta in 2000, Jeff Bezos put in $14 million in 2005, and at its peak buyers posted nearly $50 million of RFQs a day. It didn’t price a single file. Shops bid, buyers picked, same as a fax machine with a login. It ended inside Shapeways, which went Chapter 7 in July 2024. I bring it up because a lot of readers will say “we tried online quoting in 2008 and it didn’t work,” and what they tried was an RFQ board. Different thing.

What the Rules Page Actually Does

Jim Belosic of SendCutSend describes his company in one line:

We’re trying to be the In-N-Out Burger of CNC: tiny menu, fast, good, and cheap.

Here's the tiny menu in practice. SendCutSend takes DXF, DWG and STEP, no prints. The min/max table says what fits: up to 44 by 30 inches for an instant quote, a minimum part size for each thickness. A hundred-odd DFM checks run when you upload. And on tolerances, Jim's rule is the one that holds the rest up:

We treat the entire part as “critical” and the goal is to match the model 100%.

No GD&T, no “which feature matters.” The file is the spec. By his count the company has made 50 million parts with zero prints.

OSH Cut’s capabilities page is the same idea with different edges: parts have to pass the DFM check, ±0.005”, ships to the 50 states, and perfect cosmetic parts are “outside our current capabilities.” Then the rule under the rules: prices are “instant, online, and the same for everyone.” No negotiation, no net 90 for the big customer, no better price for the guy who calls you directly. Caleb told me on my podcast two years ago that OSH Cut quotes 16,000 part numbers a week with zero quoting staff.

They refuse whole markets too, not only parts. Someone priced a small widget at $11 from China against $260 from SendCutSend, and Caleb said what it meant for the next obvious lane, CNC:

This is one reason I’ve dragged my feet on adding CNC machining. If we do it, we won’t be focused on small, lightweight widgets. Going head to head with China in that space is suicidal.

Jonathan Sessa at Fabworks says metals only, and explains his refusals plainly:

Some requests are not crazy, we just are not a perfect fit, maybe we don’t have the exact right tooling etc. Streamlined systems will always win in the long run here.

Kenneth Cassel of RMFG in Fort Worth publishes his prices (from $1 to $4 a bend, taps at $8 base plus a dollar or two a tap) and says what he doesn’t have: “we don’t have a CNC mill (yet).”

twentyfour26 in San Francisco is two guys, Hemant and Sam, with their own CNC. They take one part size, 300 by 150 by 90 millimeters, and one stocked material, 6061. Anything else you have to ask for. Their site says “’No’ should still be useful.”

They all do one more thing: they only work on parts they made themselves.

SendCutSend anodizes and welds only what it cut. Xometry won’t grind or finish a part you mail in; it has to be a part that came through a Xometry order, even though a partner shop made it. MacroFab’s 10-day board assembly stops at 50 units and 2,000 placements, no through-hole parts. The reason is simple. If you don’t know what shape the part arrives in, you can’t put a price on it. So they don’t take it.

The Graveyard

Now the ones that said YES to everything.

Plethora, San Francisco, founded 2013, raised about $36 million from Founders Fund, Lux, Autodesk and others for instant-quote CNC with automated programming. Only 3-axis parts actually got programmed by the software; 5-axis and turning needed a person. It bought a 57,000 square foot building in Georgia before the orders came. The post-mortem, from investors and former employees in Not Boring:

They tried to do everything in software, cutting the humans out of the loop.

“The offering was too broad.” “Plethora wanted to do everything.” It ended in a timed online auction in February 2022.

Tempo Automation built a fast board-assembly factory in San Francisco on $74.6 million of venture money, then went public through a SPAC. Its own prospectus warned what would happen:

If Tempo is unable to utilize Tempo’s owned manufacturing facilities at a high level, the fixed costs associated with these facilities will not be fully absorbed.

In its last full quarter it kept 2.6 cents of gross profit for each dollar of sales, and spent more on overhead than it took in. It went Chapter 7 in December 2023 with almost nothing left to sell.

Fast Radius in Chicago (Mitch Free of MFG.com started the first version of it in 2014) went public at a $1.4 billion value, promising $635 million of revenue by 2025 from any process, any part. The real business kept 7.7 cents on the dollar and was sold out of bankruptcy for $15.9 million.

Shapeways is the sad one, because it started right: four materials, one file type, a price per cubic centimeter. Then it widened to any process and any customer, revenue stopped growing for five years, and it went Chapter 7 in July 2024.

The survivors paid too. Protolabs bought 3D Hubs for $280 million, a marketplace doing about $25 million a year, and less than two years later took a $118 million goodwill write-off; the Hubs deal was most of the goodwill it had.

Look at those two numbers on the card, 2.6 cents and 7.7 cents. That’s what was left after making the part, before rent, before the office, before anything. A machine shop couldn’t survive a year on that. The website promised a price, and the price was wrong, because the real cost of a 5-axis part, or a board full of parts you still have to buy, is not in the file. An instant quote is a fixed-price contract signed by a computer. If the computer can’t see the cost, you just sold insurance for free.

Xometry makes the same mistake and survives it for one reason, and the 10-K says it plainly:

From time to time, we may quote buyers prices that are lower than the cost we agree with a supplier.

A marketplace can hand a mispriced job to another shop. A factory can’t hand it to itself.

One more, the only first-hand story I found of a small shop that built a lane and let it go. Zane Hengsperger, on his father’s Detroit-area machine shop:

Hired software engineers, bought some 3D printers, and tried competing with Xometry and Protolabs... The company sold to Sybridge Technologies. Which was trying to do in 2020 what SendCutSend is really good at. I don’t know why they dont do the insta quoting anymore, but I fear it was a lack of GTM, culture, and accuracy of the product.

GTM, culture, accuracy. I’ll come back to that, because I watched it from the software side.

Hadrian Productized the Factory, Not the Part

I have to put Hadrian in the middle of this. It’s the biggest bet on the board, and it’s a different kind of bet. I don’t know the company from the inside. But what’s public is enough.

Go to the homepage and look for a part size limit, a price, a quote button or a lead time. There is none. No list of materials, no list of processes. Hadrian sells three things: parts, a dedicated cell in its plant or yours, or a whole factory. Its software, Opus, “interprets legacy designs.” In plain words, it takes your old drawings. That is the exact thing SendCutSend refuses. The job posting tells you who runs the machines: “Every technician starts at $20 regardless of experience,” trained in 30 days. Some came from nursing and retail. Some run 10 machines at once. Small customers don’t get on the floor at all; they go to a partner network. And per Chris Power, 80% of the revenue is contracts of three to seven years.

So what Hadrian put on the menu is the factory itself: the machines, the software and the training, built so a new hire can run it in a month. SendCutSend says no to the part that doesn’t fit. Hadrian says no to the customer that doesn’t fit: nothing small, nothing short, big primes only.

You know the money part already: a $7.87 billion valuation in August, a 2.2 million square foot plant going up in Alabama, and last Wednesday a seven-year deal to run cells inside Lockheed’s missile plant. What you don’t know, and I don’t either: revenue after 2024, how many parts they ship, what a part costs them, and if any one of the factories makes money. The Baffler went to the ribbon-cutting in Alabama and saw a floor that “lacked functional machinery of any kind.” Can a $20-an-hour technician with a month of training hold aerospace tolerances on ten machines, and make money doing it? That’s the question the bet rides on. No one outside the company can answer it yet. Either way, the lanes get built with or without Hadrian. The sheet metal guys already proved the model at a smaller scale, with their own cash.

Things a Website Can’t Price Yet

In July Jim Belosic of SendCutSend asked his followers which process is the biggest pain when you're getting something made. He went first: heat treating ("there's like one guy who's good and he's 6 weeks out"), plating and polishing, assembly, packaging, domestic PCBs. Dozens of replies later you have a map of the lanes that don't exist, and the replies tell you why.

The furnace. Heat treat is priced per cycle, per pound or per piece, and a commercial heat treater’s own guide says it in one line: “everyone is priced a little differently.” The price depends on what else goes in the furnace that day. There are 697 heat-treat shops left in the US, down 11% in five years. Aalberts closed Fort Smith, Wilse closed two plants in Ohio, Bodycote announced six closures in the first half of 2025. And Jim’s one guy is six weeks out.

The mask. Plating, anodize, chem film and paint: about 1,990 shops by the census count, and the finishing industry’s own report says about 1,900 finishing operations have closed since 2000. Two blockers: masking and Nadcap. One reply to Jim: “Multi step Anodizing /chem film WITH MASKING (nadcap).” Jim: “Masking is hard.” SendCutSend’s anodize line, the only bounded one I found, does Type II only, three alloys, five colors, up to 23 by 23 inches, and the FAQ says “We cannot provide masking.” Jim on what the line cost him:

It takes 5x whatever time/money budget you are planning... No discharge to sewer or atmosphere. Recycle your water. Full containment.

The die, the hob, the pattern. Gears: 205 shops, and one reply called gear cutting “the biggest shortfall in the market right now.” Castings: aluminum foundries down 20.7% in five years, and a foundry VP in Casting Source says new buyers “don’t understand why they can’t get it in four days, let alone four weeks.” Forgings: 152 plants left, more than 241 closed or consolidated in two decades, closed-die lead times of eight months to a year in one Scot Forge case, and the DoD’s own metallurgy lead saying the defense cast-and-forged base “has shrunk by 80 percent.” Extrusions: a $1,000 to $4,000 die and a minimum of 1,000 pounds or more. In each case somebody has to make something before the first part, and the file doesn’t know what.

The gauge. Grinding. “Precision grinding is a cause of no-bid syndrome,” one reply. Jason Premo of Acclaim Aerospace is looking right now for a domestic grinder for 10,000 to 100,000 bearing raceways, with a warning:

If you don’t have the metrology ability on raceways like this, no can do.

A commenter: “This job has been floating around.” The part comes in hardened with a history you don’t know, and goes out with a 5-micron form check on a gauge the shop may not own.

The BOM. PCB fabs in North America went from more than 1,500 in 2000 to 199 in 2020. The bounded version exists (MacroFab, Screaming Circuits), but the bill of materials is the variable no lane has bounded; Tempo listed component shortages in its risk factors right up to the end.

The cert. Nadcap takes six to twelve months to get and it’s per process. Customer-mandated vendors (”Customer who makes you use there preferred Plating or Heat Treating vendors,” one reply; Jim: “Real”) put somebody else’s queue inside your lead time.

And one more line from that thread, about pricing. Jim, on how the specialist tier prices:

Usually the price is “revealed” by invoice about 2 months after the job is done and you’ve moved onto the next project.

The un-productized shop doesn’t quote. It invoices. That’s the last pricing power left in the chain, and it’s why Jim’s thread is full of complaints about those shops and those shops are still there.

The Money

Is a lane worth building? Here’s what the public companies say, and I’ll keep the numbers short because the chart carries them.

Start with the two public ones. Protolabs makes 44.5 cents of gross profit on the dollar, down from 61 cents when it went public, and has grown under 3% a year since 2018. Xometry makes 34.7 cents and still loses money. Per employee, SendCutSend does about $244,000 a year, Protolabs $213,000, and the best job shops in Gardner’s survey $190,000. Better, yes. Not ten times better.

The gap is in the customer. SendCutSend’s biggest customer is 0.59% of its revenue. The job shops I see for sale have one customer at 30 to 40%, and the buyer knocks the price down for it. OSH Cut gets paid by credit card at checkout on about 95% of orders, $60,000 orders included, Caleb told me on the podcast. The job shop waits 60 days for the same money.

And speed. Caleb Chamberlain of OSH Cut wrote on his blog that rush fees were “literally half of our net income.” So in late 2024 he cut the standard lead time from five days to two. He lost 75% of his rush customers, and growth more than doubled. His reason is simple: “the bigger you get, the less excess capacity you need, and therefore the less you have to charge for speed.” A shop with one laser needs a rush fee. A shop with fourteen doesn’t.

Jim raised $110 million in May at a billion-dollar valuation, after eight years of profit on a $6 million friends-and-family round, and his reason was this:

Over the last 12 months demand has been steadily outpacing our capacity, and we’ve run over 35,000 hours of overtime to keep up.

Jonathan Sessa at Fabworks has the plainest version of the economics, replying to Caleb’s doubt that software unlocks 3 to 5x capacity:

Funny because we were right there, and did the transformation, and are LESS efficient per headcount lol. Labor is not a big slice of the COGS for a mfg this size (it’s material and buying capex), and an “at capacity” legacy shop can run wildly smooth with a small team. You just can’t scale because cashflow is so bad... we are wildly more profitable thanks to better unit economics on parts, ie having good cashflow to buy capex.

Same people, better cash, more machines.

The Lane Starts Designing the Part

Here's where I think this goes next, and the guy who sold the first sheet metal lane to Protolabs wrote it down last Wednesday. Jay Jacobs, who built RAPID and sold it to Protolabs, tells a story from around 2015. An engineer at PTC wrote a side program that changed a sheet metal design over and over inside Pro/ENGINEER and checked each version against RAPID's SolidQuote plugin for the price.

After more than 2,000 iterations, his program landed on the part design that achieved the lowest cost.

It went nowhere in 2015, because nothing around it existed. Jay’s argument is that it exists now: CAD from a prompt, instant prices from Xometry and the lanes, agents on both sides. His line for what comes next:

Going forward, design intent will drive a recursive process where manufacturing drives design.

And the missing piece, in his words: "accurate, real-time manufacturing cost tied to geometry and prints." Which is exactly what a rules page is. SendCutSend's min/max table and OSH Cut's DFM check are the first cost functions an engineer's software can call without a human. Jay asks the CAD companies a question I'd pay to hear answered:

If geometry comes from a prompt and agents run the loop, what exactly is the CAD company selling?

And he makes a call on shops that matches what I see:

Small and large shops will thrive. Medium shops without intention will struggle.

So the lanes won’t only take the parts that fit them. In a couple of years they’ll help shape the part before anyone draws it, because the design software will check each version against their prices. The engineer won’t be designing for your shop. He’ll be designing for whichever shop put its rules online.

My Take

A lane is a list of things you won’t do, and the software is there to make sure you don’t do them. The rules page is the product. The laser you can lease. Jim calls SendCutSend “more of a software company than a traditional manufacturer,” and I think what he means is this: the hard part is deciding what to say no to, and then holding that no when a big customer asks you to make an exception.

A job shop with high mix can’t squeeze itself into a website, and it shouldn’t try. Under my Stuck at $3 Million post a reader called the lanes a cop-out, because they skip the hard part of a job shop. Fair enough, as a description. But as a plan, copying SendCutSend is what Zane’s father’s shop tried, and the buyer dropped the instant quoting. What a job shop can do is the same trick on its own repeat work. Take the five part families you run each month. Write the rules for those. Quote those in minutes. Everything else stays custom. Precision Profiles did exactly that and made Top Shops.

The money I’d chase is between the lanes. Here’s how Caleb sees it, from his comment under my post, a few big productized services:

Serving a large ecosystem of more specialized, hands-on manufacturers who take advantage of the services we provide

Michael Kornuta built NORAMARK as that other half on purpose:

You won’t get instant quotes here but most jobs get at least one quote within a few hours... it’s really the best deal for anything that doesn’t fit the model of

It's for everything that doesn't fit the instant-quote model, he says. The open layer, where an engineer's CAD can see the lanes and the specialists at once, is the thing I'd put my own money on, and after Jay's piece I'm more sure of it. It doesn't exist yet.

The Rule

Here’s what has to be true before a website can price a process. Find a lane that breaks one of these and still prices instantly at scale, and the rule is wrong.

1. The file is the whole order. Outside the envelope, the shop says no instead of quoting.

2. The material is already on the rack. Zane: “there are no fast lead times without being able to touch the metal.”

3. Nothing gets made before the first part. No die, mold, hob, mask or fixture, unless the software makes it from the file and throws it away.

4. Machine time comes from the geometry. Cut length, bends, volume. No engineer deciding how to run it.

5. Inspection is cheap. The day a customer needs a 5-micron form check signed off, the lane closes.

6. No masking.

7. No outside paperwork per order. No Nadcap, no mandated vendor, no certified weld. Each one puts somebody else’s queue inside your lead time.

Two things follow. The lane only finishes what it made. And it makes money by piling small orders onto shared capacity, which is why its first customers are the ones a job shop turns away. The bigger the pile, the cheaper speed gets.

The exceptions prove it. Protomold broke number 3 by making the tool straight from the file, and it’s the lane that ran out of room first. Connor Kapoor at Digital Metal breaks it on purpose with printed trees and pays for it: “1 in 5 introduces a new edge case or problem we need to solve.” Heat treat fails two of these on paper and can still work inside a small menu, which is what Charlie Hushek is building.

Who’s Building the Next Ones

Early, I mean really early, so take these as signals. Charlie Hushek, fifth-generation heat treater at Phoenix Heat Treating, about 2,000 customers and 400 alloys, in April and then again last Tuesday:

First draft of the heat-treating instant quote tool is now in internal testing! Once this goes live it may be the first in the nation

I think we are on our way to being a @sendcutsend equivalent for heat treating.

No launch yet. Connor Kapoor at Digital Metal in Austin: a 400 square foot garage in March 2025, a factory four months later, “$3M+ in quoted volume,” an A356 manifold at $258 with a two-day lead, aluminum first, now 17-4PH, and the quote engine sold as an API that other instant-quote shops call. David Liu at finalREV in Berkeley, who ran in-house manufacturing at Neuralink before this: STEP files only, 3-axis and small mill-turn up to an inch round, same day in the Bay Area, two days anywhere else. Kenneth Cassel at RMFG: $4 million raised, 2 to 9 people in a year, a tube laser “running ~16hrs a day,” and a hosted MCP where an agent can place a real order. Zane Hengsperger at Nox Metals, one step upstream of the rest, raw metal with certs in 24 hours: “Not Saas. We don’t sell software. We sell metal.” And Kenneth’s name for the cohort, when a thermoforming instant-quote site launched last month: “The instant quote cinematic universe expands!”

Caleb’s advice to anyone reading this with a shop and an itch:

If you are looking for an opportunity, just grab something that doesn’t have a productized lane yet.

The list is in the section above on things a website can’t price.

Builder Read

The lanes are being built by factories with their own cash, in Reno, Spanish Fork, Ventura, Fort Worth, Austin, Berkeley and Phoenix, and Caleb’s “decades-long grind” is the right expectation. Don’t build another instant-quote tool; three of these guys built theirs in a few months. Build one of the two things around the lanes. One is a rules engine for a job shop’s own repeat work, inside the ERP it already has, so what Precision Profiles did by hand becomes a product. The other is the layer between an engineer’s CAD and all the shops out there, lanes and specialists both. Caleb calls it MCP, Jay Jacobs calls it the facilitator. The lanes will publish their rules. Somebody has to read them for the 16,000 shops that won’t.

I’m Watching This

• Phoenix Heat Treating’s instant quote tool. When it launches, the test is in The Rule: which alloys, what weight limit, what it does with a Nadcap job.

• Fabworks’ assembly welding (”coming soon”) and SendCutSend’s multi-body weldments, still through sales.

• Digital Metal’s steel castings, 17-4PH from a STEP file. The 1-in-5 edge-case rate is the number to watch.

• Jay Jacobs’s five deep dives on the design loop; he promised one per shift.

• Caleb’s prediction: from four lanes to a dozen in a few years. Check the pricing pages in 2027.

Const

I spent last week in Cleveland at Top Shops and missed my usual Tuesday newsletter for it. It was worth it, and you’ll see why.

There were two different conversations going on in parallel, at the conference and online, about the same thing: 16,876 machine shops. If you had told me five years ago that one of the biggest VC funds in America would be posting about small machine shops, I wouldn’t have believed you. But here we go. This is the a16z post that got shared hundreds, if not thousands, of times:

Nothing in it was new to me. After 12 years of serving these shops, neither the size of the market nor the size of the shops the whole American industrial base relies on surprised me. It surprised pretty much all of X though. Anyways.

On the stage and in the hallways at Top Shops, owners talked about setup sheets, the one guy at the shop who knows where everything is, net 90 terms, and a shop that sold for pennies. Online, a16z called small shops “friction,” and Caleb Chamberlain of OSH Cut wrote that MCP servers and AI agents mean fewer shops, at billion-dollar scale. I was torn between the two. But I think they’re talking about the same thing. Let me explain.

This issue is my read of where the US machining market is in October 2026: what holds the small shop back, where it’s going, and why I’m bullish anyway.

Same Machines, Four Fewer Hours

Gardner's Top Shops benchmarking, presented in Cleveland on Tuesday (top 20% of respondents against the rest; the full results aren't online yet, I'm working from the slides): a Top Shop runs 15.5 CNC machines against 15 for the rest. Same machines. It runs them 12.75 hours a day against 8.8, and makes $353,000 per machine against $180,000. Net margin 16.5% against 8%.

Now, here is the number that bothered me, and no one on the stage talked about it. Sales per machine is flat since 2018 for both groups: $359,000 then, $353,000 now for the Top Shops; $176,000 then, $180,000 now for the rest. Eight years of "reshoring boom" and the median machine makes the same dollars it made in 2018. Less, after inflation. The Top Shops' margin went from 10% to 16.5% in the same eight years. So they made more money from pricing and from picking customers, not from more parts through the spindle.

Xometry's 2027 outlook, out October 1, says the same thing from the other side: 44% of manufacturers under $50 million in revenue run below 70% capacity, against 16% of those above $100 million. Among the shops with idle machines, 48% say their biggest obstacle to new business is finding enough work. Stacy Sutton, a fractional CMO, put it this way on LinkedIn:

So the machines are there, the people (mostly) are there, and the work isn’t coming in. That’s a go-to-market problem wearing a capacity costume.

She sells positioning, so take the framing with that in mind. But the number under it is Xometry’s survey, and it matches what I heard on the floor at Top Shops.

The Machine Shop Is the Bank

The four honorees this year, Astro Machine Works, HEI Precision, Hill Manufacturing and Precision Profiles, said what I'd call the same thing in four different ways: no flat shop rate, price by industry, know what the customer costs you before you price him, and net 30, 60, 90 goes into the quote or you don't quote.

Mike Payne of Hill Manufacturing, who has bought and run six shops of this size, said it plainly. You take the big customer at net 90, and:

You might be breaking even at the end of the year because you had to borrow on a line of credit to cover their payment terms.

Hill Manufacturing went from 90% oil and gas to 60% in eight years, on purpose. Yep, diversification.

The owner is a humble machinist who banks his OEM without being a bank. He floats 30 days from PO to invoice, then net 60 on top.

Jason Premo of Acclaim Aerospace wrote it out under Colin’s thread that went viral:

He pays the NRE out of pocket until parts ship. He has no banker who understands the business (Colin's line in the same thread). So he can't say no to the wrong customer, and he can't quote fast enough for the right one. Gardner's quoting numbers are the proof: Top Shops send 70 new-part quotes a month against 25, answer half of them inside 24 hours against 29%, and win fewer of them, 41% against 45%. They quote more, faster, and they can afford to lose.

The difference between the two groups is the terms and the quoting. The machines are the same. Bottom line, balance sheet, cash flow first. Everything else after. This also answers why so many shops ignore automation, robots, software and the rest: there’s no cash (and time) left for it.

Chuck, and the Shop That Sold for Pennies

Wednesday was the knowledge day, three talks in a row on the same problem. Paul Van Metre, who co-founded ProShop and ran a shop before that, had the survey: of 144 precision shops, 52% would degrade within a week if one person wasn't there.

"Everyone has a Chuck," Paul said. Chuck is the guy who knows where every fixture is, which drawer the special cutter lives in, and which machine needs a kick in the morning to hold tolerance. 70% of owners carry the shop's knowledge personally. The average skilled machinist is 56.

Then the slide I photographed twice. Shops at the top digitization level: 81% can run without the key leader. One level down, the shops with scanned PDFs and travelers on a tablet: 14%. Paper only: zero. The jump is from the scanned-PDF shop to the native digital one, and it's six to one.

Paul’s story from near Boston: a guy went to buy a used Bridgeport off Facebook Marketplace and found a full shop, three or four CNCs, with an owner who had run it for 37 years and had to be out by month end with no buyer for the business. He bought the rest for pennies on the dollar. The knowledge was in the owner, so the business was worth the scrap value of the machines.

James Marzilli of Marzilli Machine went paperless in 2020, bought machine monitoring, connected it to ProShop, and then hit a problem. More data stopped helping. His words from the stage:

You have to hire somebody now to kind of parse through all this information.

His best people quietly stopped trusting the system. He said they had moved the paper mess into the computer, where it’s out of sight. His fix is an agent that pulls from the ERP and the machines and does things, with him approving what goes into the company’s knowledge base for about five minutes a day. Mike Payne has one at Hill Manufacturing that reads incoming quote packages and flags the red flags. Russell Winter built a quoting tool for his shop in two days and $250 in credits, trained on his own quote history. Does it mean every machine shop owner has to vibe-code their own software? Nope. But what’s table stakes now is adopting the software (inclding LLMs) and connecting it to everything that touches your cash flow.

One more Gardner number fits here. Operators make $25 an hour at both groups. The only wage gap is the CAM programmer, $40 against $35. That’s the person who writes the process down, and the best shops pay him more. Which tells you something: if writing the setup down takes an hour, Chuck is not going to do it.

Paul showed what works in legacy shops: an apprentice walks behind Chuck and takes notes, Chuck dictates instead of writing, somebody takes photos at the machine, or puts on Meta glasses for the setup. One shop put a $500 bounty on the best-documented job and got enough setup sheets out of it to start a second shift.

Twenty Years of Productized Shops

“Productized” is not new. Protolabs started in 1999 and did $533 million last year. SendCutSend and OSH Cut both started around 2018. Fabworks, 2023, a 42-year-old Ventura shop relaunched as instant-quote laser cutting. RMFG in Fort Worth, 2024, $4.5 million pre-seed, nine people. finalREV in Berkeley. twentyfour26, two people in San Francisco, 2026. Each one did the same two things: they cut the list of what they make down to what they can price with software, and they put the rules on the website so the customer sees them before he uploads a part.

Caleb of OSH Cut said on my pod Machining Beyond the Craft back in 2024 that OSH Cut quotes about 16,000 part numbers a week with zero quoting staff and about 60 people, that 90% of the parts don’t come back, and that they don’t answer manual RFQs:

If we’re responding manually to quote requests, we’re just like every other shop.

His model is sheet metal. CNC with tight tolerances and an assembly on top is a different job, and Caleb said so himself last week: it needs “a dozen additional lanes,” one for each kind of work. More on that below.

The marketplaces are the same age. Xometry, Fictiv and 3D Hubs (now Protolabs Network) started the same year, 2013. Xometry did about $687 million last year across 4,996 active suppliers. Fictiv sold to MISUMI for $350 million, 3D Hubs to Protolabs for $280 million, Plethora closed. Twelve years in, here’s what a marketplace that works looks like: it owns the customer, takes the order, sends the hard part to a shop in its network at the marketplace’s price, and puts its own name on the invoice. The shop is the back end. Alex Huckstepp, who co-founded Uptool (AI quoting for small shops), wrote that the answer is Toast and DoorDash for machine shops. I agree with the Toast half. But look at what DoorDash did to restaurants: it owns the customer, takes a cut of each order, and the restaurant became a kitchen in the back. Not one owner in Cleveland asked for that.

The legacy shop has its own way to productize. The owners just don’t call it that. John Carpenter of Accelerant said a job shop really runs on about 50 standard setups (some shops keep 150). His advice: when a new job comes in, match it to one of the setups you already have, do the same with tooling, and push your sales guys toward work that fits those setups. The jobs look different from the outside. Inside, it’s the same setup over and over. Precision Profiles, 34 people in Titusville, Pennsylvania, does a setup in about 10 minutes (the average is 90) and quotes a part they’ve made before in 10 minutes. Their website says they’ll quote anything. Inside, the list is short.

Future and the Room’s

Caleb Chamberlain's post last Tuesday, 83,000 views: MCP servers become as common as web pages, each shop has an API with its own DFM, prices and lead times, agents design and source by searching shops. Then:

Because all of the above requires savvy most legacy shops don’t have, they either fade away or pivot. There will be fewer shops, and they’ll be at billion-dollar scale.

And no brokers, no marketplaces, no shop-to-shop tools, “not even a little.” People under the post asked the practical questions: who eats the remake when the server says five days and the parts come three weeks late, how this works under CMMC level 2, and where the record of what a shop actually shipped would live. Caleb’s answer to the last one: “I wonder where that kind of performance history could live.” And John Carpenter said on Wednesday that regulators are moving to classify G-code as CUI. If that lands, the agent needs a login before it reads your part program.

Then Caleb wrote a second post on Wednesday, and this one I agree with almost completely:

Under the hood, the buildout is less hyped and sexy. It requires manufacturers that can crank out parts, and that’s very much NOT an AI play. At least, not yet. [...] We need dozens of productized manufacturing “lanes.” [...] It’s a decades-long grind. The AI-friendly API layer is a tiny and comparatively easy component of what makes the whole engine function.

Put the two posts together and what he's describing is dozens, maybe hundreds, of OSH Cuts, built by operators over decades. Not one Hadrian. And here's what I'd add from the room: a "lane" is just a shop that wrote its rules down. Caleb says legacy shops don't have the savvy for this. He's right today, because 70% of owners carry the rules in their heads. James Marzilli and Mike Payne show it doesn't have to stay that way.

The Buyer Changed First

Austin Vernon quoted Caleb’s post on Wednesday morning:

A lot of folks are still making the assumption that there will be tens of thousands of small manufacturing shops. But that is completely unreconcilable with the reality that once a company in a certain niche solves the software backbone they can add (financed) incremental machines for a fraction of the cost of what it takes to semi digitize and upgrade a legacy shop.

Caleb’s reply: “It’s true, unfortunately.”

I know what a lot of owners say to that, because I’ve heard it in shops for twelve years. It goes like this: we’ve had these customers for twenty years, they’re not going anywhere, we don’t need the internet to find work. I understand it. The relationship pays the bills, right up until the day it doesn’t. And finding a new customer is already the hardest thing a small shop does (Xometry’s survey again: 48% of the idle shops say so).

The problem with “our customers aren’t going anywhere” is that the customer is the one moving. Xometry did $229 million in the second quarter, up 41%, with 89,557 active buyers, up 20%, and raised its growth outlook for the year to 33 to 34%. On that guidance it lands somewhere around $900 million this year (my arithmetic from the 2025 number, not theirs). Sanjeev Sahni, the CEO, said on the call where the money comes from:

What we are seeing is a generational shift in who is making manufacturing purchasing decisions. The engineers, procurement buyers and supply chain lead roles are now full of dynamic digitally native individuals. They expect the same frictionless experience at work that they have in their personal lives.

That's the generation change, and it's the part the owner at the bar misses. The engineer who used to call the shop retired. The one who replaced him orders a prototype from his laptop, with a budget he can spend without going through procurement, a buyer or accounting, and he pushes his company to let him work that way. I asked Caleb under his post who the customer is for this. His answer:

I agree with that order. And that order doesn’t reach the shop that is waiting for the phone call. The owner doesn’t even find out he lost it.

So both sides are changing at once, the buyer and the manufacturer, and the buyer is ahead. That is a much bigger change than one shop’s MCP, and I think (hope?) what Caleb described is real within a few years, not a decade.

And where the US beats China, in my opinion, is not on the floor. In Shenzhen the anodizer, the lathe shop and the engineer are in the same neighborhood. We don’t have that here, and no software fixes geography. We can win on what’s in the wire: the intelligence inside the shop (James Marzilli’s agent, Mike Payne’s quote reader, Russell Winter’s $250 tool), a tight community that passes work around (Alex Huckstepp’s line, and the one owners nod at: shops “routinely and quietly pass each other components of a BOM for which they lack tooling”), and ordering that runs through agents and CAD instead of email and a two-week quote. That needs a common language between an engineer in Austin and a shop in Ohio, with the knowledge written down on both sides. Caleb calls it MCP. The owners in Cleveland call it a setup sheet. It’s the same information, but only one copy is written for a machine.

The Money Above the Shops

One more number, because the 2x orders depend on it. Venture investors put $49 billion into defense startups in 2025, up from $27 billion. The NatSec100, the hundred biggest venture-backed defense companies, raised $39.6 billion last year and received $4.3 billion in federal awards. Eight percent of those awards are production contracts. The rest pays for prototypes and pilots. Abhishek J. laid the numbers out on LinkedIn, and Mitch Free, who founded MFG.com and ZYCI and was a machinist before that, reposted it with one line:

Throwing capital at a solution doesn’t necessarily equate to creating demand for the solution.

Elisabeth Smith of Acutec said the other half from the stage at Top Shops:

A lot of small shops can’t get anybody’s attention at some of these automation places, tool places. Because honestly a lot of these venture capital backed companies are just sucking all of the resources of machine tools, of tooling... someday they’ll need to have a return on that investment to those venture capitalist people and then... we’ll have a reckoning.

Against that: 77% of the Xometry respondents report bigger backlogs than a year ago, Paul Van Metre called it a super cycle from the stage, and Practical Machinist's survey of 577 shops last week had 42% growing and 33% shrinking. The Fed's August numbers show where the work is: defense and space equipment up 6.1% over a year, aerospace and other transportation equipment up 4.2%, consumer goods down 1.1%, motor vehicles down 1%. And the machinery companies in the September ISM report said higher rates are slowing new construction. So if you make parts for aerospace and defense, you're busy. If you make parts for consumer products or construction, you're not. That's why the median machine is flat while the Top Shops grow.

Why I’m Bullish

Because the fix is cheap and the owners on stage already did it. Pick the industry and price it. Check the terms before the quote. Quote everything, fast, from your own job history. Write the setup down at the machine, not at the end of the shift. Run two more hours a day on the machines you own (Matt Sand of Avalon Manufacturing did the math on Wednesday: two hours times 250 days is 500 spindle hours a year per machine, and that's the gap between a Top Shop and a normal one). None of that is a robot cell. Be lean, be picky, productize what repeats, and run it as a business.

And because the buyer already moved. The young engineer with a laptop and a budget is the customer the new lanes are being built for. The small shops that write their rules down now will be the ones feeding those lanes. And the piece in between, the one that connects the engineer to the shop that can actually make his part, isn't built yet. That's a lot of work, and that's the opportunity.

Builder Read

If you own the shop. This week: terms on your top five customers, real days to cash, next to the margin on their jobs. This month: a 24-hour quote standard, estimate pulled from the ERP. This quarter: the ten jobs that repeat, documented at the machine, and one agent that reads your quote packages before you do.

If you build new software. Remember the owner doesn’t have a spare hour. If your tool takes a week to set up, he’ll pick the $250 one he can build himself over a weekend.

I’m Watching This

• Caleb’s MCP demo. He asked Claude to design a firepit housing and order it through OSH Cut’s MCP. The first design failed DFM, the API told it why, and it fixed it. Then he posted that design-by-chat with parts in 24 hours is already possible.

• Gardner Business Index. 53.5 in September, ninth month of expansion, fifth month in a row of cooling. If October goes under 50, the super cycle talk from the stage ends.

• Practical Machinist’s survey. 577 shops, 42% grew, 33% shrank, and 87% of the pessimists still plan to buy equipment. Worried and buying at the same time.

Const

I’m Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I’ve seen every pitch this industry makes. Someone’s always selling hopium. I find it.

Last week we lost a customer we’d had for five years (if someone not aware, my company sells quoting software to shops). I asked what we did wrong. The answer: “No we love your app, actually!!!”. For years they had been bending our quoting technology to price a process that we never truly built it for, changing settings until the numbers matched their costs. So recently they wrote their own quoting system for that process and left. And they left because they got good enough at their own data to not need us for that part.

Same week, Caleb Chamberlain, who runs a sheet metal shop in Utah, posted an email from OpenAI. Someone from their Research Data Partnerships team had been “looking at how you combine sheet-metal manufacturing with your own design-checking software” and asked whether he had “existing designs or project examples” he could make available “for research and model development,” with a mutual confidentiality agreement attached. Caleb’s answer:

Now put the two together. A shop’s job data is worth money now, to three different people. To the shop itself, because it can learn from it and quote better next time. To the owner, because a potential buyer pays for a shop that runs on a system and does not pay much for a shop that runs on his memory. And now to OpenAI… which is emailing sheet metal shops to ask for it. Modern shops like OSH Cut, SendCutSend, Fabworks and marketplaces like Xometry, Fictiv all learn from every part they make or sell. But most legacy shops don’t. And that is what makes the difference when it comes to growing, competing, getting financing, or selling the business at the end of the day.

Xometry Learns From Every Part

Xometry has a new CEO. Sanjeev Sahni took over on July 1 after founder Randy Altschuler moved to executive chair. On his first earnings call as CEO in August, an analyst asked whether the new pricing model was a step change. His answer:

Every part that gets produced helps make the next quote stronger, which means that the feedback mechanisms on the actual production floor are actually making their way back to when the next part is priced on the site. So we already know if we’ve got a part that actually was more complicated than anticipated, what should that do to the pricing of the next part.

The 10-K says the model “continuously learns from hundreds of millions of data inputs.” Back in 2022 Randy told JPMorgan they were “retraining our data set every week.” In July they reported the model now predicts CNC cost about 15% better than before.

Here is how it works on the shop side, because most owners only see the buyer side. A buyer uploads a CAD file and gets an instant price from the model. The same model then sets a second number: what Xometry will pay the shop that makes it. The shop does not quote. The job shows up on its job board with the price already on it, and the shop accepts with one click, passes, or sends a counteroffer. Which shops see the job first depends on what Sanjeev called a partner suitability score, built from “machine capabilities, quality and on-time shipping history.” The new Workcenter puts your quality and timeliness scores on the home page, and it runs on “the same matching and pricing engine that powers the rest of Xometry’s AI-native marketplace.”

The payout is not fixed. In Q2 Xometry launched what Sanjeev called an adaptive sourcing model that will “price jobs dynamically for our suppliers.” Xometry also holds a patent from 2017 on predicting the probability that a supplier accepts a job “at a given quote.” So every time a shop accepts a job or passes on one, the model learns a little more about the lowest price that shop will take. Xometry keeps the difference between the two numbers. Its marketplace gross margin was 34.7% last quarter, with revenue up 41%.

I wrote about this on X last week and the replies from shop owners filled in the rest. Caleb said Xometry also runs low “hook” pricing to land new customers and may bid a job at a loss to win a first order, and that it is an algorithm making that call, not a person. Ken Spaulding said the price on the partner side is not even consistent for the same job. Jason Premo of Acclaim Aerospace said he took Xometry jobs during covid at roughly the cost of the raw material, because the small quantities fit in the bar ends he already had. Three shop owners, three different experiences. So if you run your part through Xometry’s instant quote to check your own price, remember that number has Xometry’s margin in it and can be different next week. Price from your own machines, your own setup and your own rate.

Legacy Shops Learn Two Weeks Later

Mike Payne runs Hill Manufacturing & Fabrication and hosts the Buy the Numbers podcast. After IMTS he wrote about how most shops learn from a job:

The part ships. Two weeks later somebody runs a job cost. Maybe they catch that cycle time ran 20% over quote. Maybe they remember to fix it before the repeat order hits the floor. Maybe they don’t.

He has spent nearly nine years connecting the pieces in his own shop and says it “still has holes.” His guess for when the systems talk to each other on their own is 2028.

One shop owner on X described his quoting for the last five years:

The survey numbers behind that come from vendors, so read them that way: Galorath, which sells estimating software, found that 96% of the 180 manufacturers it asked still estimate cost in Excel.

The network is not the only way to close the loop, and the software side of this is crowded now. Quoting tools like Paperless Parts, AMFG, Uptool and DigiFabster (mine) sell the shop a way to price from its own history. Toolpath and CloudNC come at it from the CAM side, reading the geometry and giving you a cycle time before anyone has cut a chip. ERPs like ProShop, Fulcrum and JobBOSS² are where the actuals land after the job runs. And two newer companies, WarpCore, built by James Marzilli, who owns Marzilli Machine, and Machine Research, sell a model trained on one shop only, with the actual time and cost of every finished job going back into that shop's next quote.

The Shop Isn’t Worth Much Without It

A fellow entrepreneur who buys legacy shops wrote to me after last week’s issue. He has bought a few shops in the Midwest in the last two years, buildings included, and brokers now send him three to five more shops every month. And most of them from owners near retirement. His description of a typical one: nothing invested in 15 years, a team with an average age around 60, a shop that isn’t worth much on its own. The owner’s retirement plan is the building, because industrial real estate went up a lot since he bought it. The quoting, the routings and the customer history are in the owner’s head, so they don’t transfer. He also said money is the problem on his side too. Banks don’t like these deals, and the family offices that call him have money but no young operator to run a shop.

A comment from Greg C. from the last week summarizes it well:

A shop where quoting lives in the owner’s head is worth less than one where it lives in a system, because a buyer is buying the system rather than the man.

Now look at the other end of the same market. Jim Belosic posted last Wednesday that SendCutSend is at 700 employees. In May he raised $110 million from Sequoia and Paradigm, and he said what it was for:

The capital is going to be used for stuff I can’t get a loan on: tripling the size of my software team, hiring two or three hundred people.

This is not only a shop problem. Greg Isenberg posted a guide this week on buying businesses from retiring owners and running the back office with AI agents. The money behind it is real: Thrive Holdings, which OpenAI owns a piece of, already runs more than 50 accounting firms and is putting $1 billion into buying more, and General Catalyst has $1.5 billion for the same play.

One step in Greg’s guide is worth an owner’s attention. Before he buys anything, he asks the seller for 20 to 50 finished jobs and goes through each one, to see what the work really looks like.

Read that from the owner’s side. The buyer is paying for the customers who trust you, the certs, the approved vendor status, the machines and fixtures, the crew that knows them, and the 30 years in your head and your foreman’s head. A startup has none of it on day one. The buyer pays less for what’s only in your head, because he can’t be sure it stays when you leave.

The rest of the guide, buy a business and triple its profit with agents built in the evenings, I don’t buy, and Mark Ajzenstadt said why best:

The buyer is learning to run a company and rebuild how it works at the same time, with payroll and loan payments running the whole time.

Manufacturing is not on Greg’s list because his profit comes from replacing people with software, and in a shop the work is done by a machine and a human. The office side of a shop is on his list.

Two more things from his guide that matter if you own a shop. His buyers pay 60 to 70% at closing and the rest later, and they ask the old owner to stay for the handoff. That’s the buyer telling you the shop is worth less without you. And his first 60 days after buying are spent sitting with the senior people and writing down what they know. You can do that now, while you still own the place. Get it out of heads and into a system, the quoting, the routings, the job costs, the way you handle each customer, and the price of the shop goes up with it.

Where I Land

Here is where I land on all this. Xometry wants to know what price your shop will take and if you ship on time. OpenAI wants your designs and data. A buyer wants a shop he can run without you in the building. Three different people, and they all want the same thing from you: the history of what you made, what it really cost you and what you sold it for. Xometry has been collecting that from its network for ten years. Most legacy shops still keep it in the owner’s head and one spreadsheet.

Last week’s issue got 108,000 impressions on LinkedIn and the comments turned into a fight about pay. (Some people read the title as the owner taking home $15 million a year. It was the shop’s revenue. I fixed the title, my mistake.) But the people in that thread who actually run shops or buy them were not arguing about pay. They were talking about the system, the loop, the record. And this is exactly what I wanted readers to take away.

Builder Read

If you own a shop. Check your estimate against the real cost the week the part ships, not at the end of the month, and keep that number where you can see it when you quote the next one. Buy the software or build it yourself, it doesn’t matter, as long as somebody else could run it if you're out for a month. And when a platform or an AI lab asks for your data, either give them a price or tell them no, like OSH Cut told OpenAI. Don’t give it away for a “partnership.”

If you buy shops. Before you ask for anything else, ask for the estimate and the actual cost on the last hundred jobs. If the answer is “ask the owner,” you are buying a customer list and a building. Pay for a customer list and a building.

If you build software for shops. A model trained on one shop is your answer to Xometry’s model, and it’s a much easier thing to sell to an owner than “we learn from everyone.” The shop keeps its data, you keep the model. My guess, and I’m not sure about it: within two years, when a shop gets sold, there will be a separate line in the price for the quoting history, the same way software companies get priced on their data today.

I’m Watching This

• Xometry’s supplier payouts. Sanjeev said the new model prices jobs “dynamically for our suppliers.” Neither the call nor the release says whether shops are getting paid more or less. Q3 numbers come in November.

• SendCutSend’s next five states. 350 employees in May, 700 now, new sites announced for Pennsylvania, Ohio, Indiana, Las Vegas and Atlanta. One reply to Jim’s post: “We’re too poor for them, so we turn our eyes to the east. I’m no Raytheon.” Every new SendCutSend site is built to win that order back.

• Top Shops in Cleveland, October 6 and 7. Gardner’s benchmark has the shop-side numbers on job costing that I could not find anywhere public. I’ll be there.

Const

I’m Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I’ve seen every pitch this industry makes. Someone’s always selling hopium. I find it.

I talk to factory owners a lot. Young hardware companies, family metal shops that started way before I was born, every background and age I can find. Everyone has their own truth about why they succeeded or failed, and sometimes things outside your control just won’t let you get where you deserve to be. But the more I dig into the physical world of factories (12 years of it now), the more I see that great engineering and machining skill now has to come with even greater storytelling, or you don’t get the attention, the talent, or the money.

For a few decades software engineers had the best paid jobs and everyone outside that circle accepted mediocre pay. A blue collar worker stayed a blue collar worker no matter how good he got, with no way to get paid more for his skill. Things changed. Once the country noticed how much manufacturing it had outsourced, machinists and welders got the opportunity of a lifetime: keep building stuff and also get funding, attention and talent. IF they can tell the story, and are ready to jump into the massacre of entrepreneurship. The irony is they also got the fruits of the software era, LLMs that let you code your own machine shop without a decade of software engineering behind you. Two worlds collided, in a good way. This story is still being written, and early success does not mean they make it in the end. But the opportunity is there for those bold enough to take it.

Two guys from the same kind of floor show what I mean.

Doug Rizzo is 62 and has spent 55 years in machine shops. He started on drill presses in his father's shop at six. His father took $400 of savings and, in Doug's words on Swarfcast this month, "turned it into a $15 million a year empire." Doug worked there 30 years and assumed he would take it over. The shop closed in 2009. Doug bought two of its CNC lathes for a few dollars so his father's partner would not be short, then spent three years in other people's shops running Swiss machines and boring mills before opening his own. Now he runs a plant during the day for the salary and the insurance, and his six machines at night for what he calls the gravy. Noah Graff asked him why. "I'm supposed to be there."

Zane Hengsperger is 26 and started scrubbing his father’s shop floor at 14. That shop is gone too. In 2017, two years before it sold, his dad hired three software engineers to build instant quoting: “This Xometry and Protolabs thing, I’m going to compete with them.” Zane went to Uber, came back, bought a fab shop with friends, and drove deliveries to 15 shops a week for a whole summer. “No screens. It felt like 2002 to me.” Then Y Combinator called him, the batch started the next day, and he flew out with “not a line of code, not a factory, not a saw.” NOX raised an $11.5 million seed round in June, on top of earlier money, and the forklift driver gets equity after three months.

You see, similar childhood and same floor. One is 62 and works two jobs, by choice. The other is 26 with a seed round and a growing team. Both of them jumped. Doug bought two lathes and opened his own shop on his own money. Zane flew to San Francisco with nothing and asked strangers for millions. I believe the difference is usually how far you are willing to jump into uncertainty, with no guarantee it ends well, and whose money you jump with. The guy who jumps furthest is the one who ends up in front of somebody with a checkbook and a story to tell.

Four Ways A Machinist Gets Paid

Start with the wage. BLS says a machinist makes $58,750 a year, median. That is the first way, and most machinists finish right there. Doug's father built his $15 million a year with Acmes and Wickmans running 12 hours at a clip on long-run work. He bought the machines, found the customers, took the hit when a run went bad, and sold parts by the thousand. Every shop owner lives on that margin, and a wage never includes it.

The second way is owning the shop yourself. Russell Winter, 32, bought Center Tool from his father in 2019 and publishes his own numbers: 6.8% net. He also tried raising money to buy shops like his, and changed his mind:

If they’re old and have depreciated equipment and old buildings and they’re making an 8% margin at the end of the year, it’s not an attractive investment.

On BizBuySell, machine shops sold between 2021 and 2025 at a median price of $840,000, after 206 days on the market. A house, basically (in most of the country anyway). And those are the ones that found a buyer. McKinsey counted 510,000 small businesses that left the market in 2022. 92% just closed. 5% sold.

The third way is to stay a welder, but pick a company with a huge story attached. Juan Hernandez started at SpaceX in 2015 at $28 an hour and they gave him $10,000 in stock. “It wasn’t a big deal. I didn’t know anything about it then.” At the IPO price it was worth $880,000. By the close of the first day, $1,046,175. Eighteen years of a machinist’s pay in one morning. A launch engineer at the same company walked away with $13.5 million. And the million was on paper. By July 10 the stock had dropped to $145 and Juan was still locked up, not allowed to sell a single share. There is also exactly one SpaceX. A welder who took the same $10,000 in stock at almost any other hardware startup in 2015 has zero today, because most of those companies are gone.

The fourth way is Zane. Also Kenneth Cassel at RMFG, who did gas station maintenance 50 hours a week, taught himself to code on a Raspberry Pi, and later put $200,000 of his last $300,000 into a fiber laser. Also David Liu of finalREV, who ran a shop in San Francisco and then set up manufacturing inside Neuralink. Julian Fried, a welding foreman from eastern Pennsylvania, now building robotic welding cells at Advanced Metal Research next to SpaceX in Hawthorne. And Jonathan Sessa, third generation at Fabworks in Ventura, who took a fab shop his family started in 1980 as a manual weld-and-cut operation and turned it into a laser floor running on software they wrote themselves. None of them is only a machinist or a welder. All of them are the trade plus code. Four of them took money that comes with a clock. Jonathan did it inside the family shop.

Doug's father answered to his customers for 48 years. These founders answer to a cap table, and they picked that on purpose, because it is the only money that funds the biggest version of the idea. And behind them I know a few more who are not public yet. Welders and machinists writing their own shop software, building parts for robots, solving a problem they used to have on their own floor. Some raised a little, some are doing it on their own savings, which is the harder way. It is the riskiest path in this industry and the one with the biggest upside, and every one of them chose it.

Nobody Writes An $11.5 Million Check For A Job Shop

Doug's father started with $400 and used multi-spindles from Graff-Pinkert. There were 14,469 banks in this country in 1984 and 4,135 by 2022. A bank will still lend you enough for a used Mori, but nobody writes an $11.5 million check for a job shop. Jeff Greason explained why last week:

“If you can’t tell the story of how you might be a 100x unicorn you can barely play.”

So the guy who gets that money is the guy who can write the software half himself. Zane's father saw it coming in 2017 and paid three engineers to build it. Nine years later Zane and his team wrote their own in a weekend. Zane posted in March:

If you work in a factory, get cracked at claude code, you have a baked in tribal knowledge moat that 96% of people don’t have.

The software his team wrote that weekend would have taken four months and three people before AI coding tools. And his investors counted fewer than 40 software engineers across all 6,300 metal service centers in the country.

What a legacy owner knows lives in his head, and a buyer cannot purchase that. Whoever buys Zane's company gets the software, and it keeps running the place after he leaves (if ever).

Where I Land

Mike Wagner of Cornfield Customs had a plan. Get very good at TIG welding, get a job at GE or in aerospace, sit in a clean room and weld parts for good money. He said it on It’s All About the Build last month:

“And then once I graduated and got a job, and I realized I don’t like people telling me what to do with my life, that’s where it ended, and I opened my own thing.”

Doug’s father made the same call in the early 1960s, with $400 in savings.

Welders and machinists are now the hardest people for a hardware company to find. The software half you can start for almost nothing. That is why 66% of the shops in this year’s reshoring survey say hiring a welder or a machinist is very difficult or already a crisis. The skill was always worth more than the wage. The guy who had the skill never had the money to own a business built on it. Now he can get that money. But only if he can tell the story himself.

The best argument against me came from Nick Huber on The Sweaty Startup in July:

“Let me just warn folks that these old boomers know how to run businesses a little better than you think.”

His guest added:

“They are savages, man. They will take your lunch money.”

Fair. Doug's father built $15 million a year out of $400. The old guys are not the problem. They knew the trade and they had a bank that would lend against a lathe. That bank is gone. What's left is money that wants a story, and for the first time a guy with good hands can write his own. The savages built their empires with the tools of their time. These guys are doing the same thing with the tools of theirs.

Builder Read

If you have the hands. Most guys stop at the wage. Owning a shop is a good living, and if you ever find a buyer, it pays for a house. Stock at a company with a big story is worth nothing most of the time, and once in a decade it is worth eighteen years of pay in one day, so take it when somebody offers it. The fourth way is the only one where you own the upside on your own work. Know what that costs. Every round you raise sells a piece of the company, and by the exit your piece can be small. Jonathan’s way, inside the shop on your own money, is slower and you keep all of it. Caleb Chamberlain built OSH Cut the same way, from a quoting app he wrote himself in 2018 to one of the biggest laser shops in the country, with no investor on the cap table. Zane’s way is faster and you keep a smaller piece of what could be a much bigger company.

Venture money wants a story about growth, and in a shop that story is almost always the software, because the software is what grows without buying another machine. Zane did not raise $11.5 million on a metal-cutting business. He raised it on one that runs on code his team wrote. A bank will still lend you money for a lathe. Nobody writes a big check for the hands alone. Learn the software half, or find a partner who has it.

If you own the legacy shop. Your buyer is somebody with an SBA loan paying about three times what you take home a year, or a 26-year-old who walked through your building and saw 2002. The difference between those two prices is whether the shop runs on what is in your head or on software that stays after you leave. Start putting it in now. Buy it, write it, or both, as long as the jobs, the routings and the prices end up in a system a stranger could run. It takes years, the first version will be bad, and you pay for it out of the shop's own money. But it has to be yours, and it has to start.

Const

I’m Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I’ve seen every pitch this industry makes. Someone’s always selling hopium. I find it.

Of the manufacturers who brought work back to the US, 65% say they are satisfied with the result. A year ago it was 96%. The share who say they are dissatisfied went from 4% to 25%.

Those numbers are in the 2026 USA Reshoring Survey, Exhibit 5, page 12. The survey is run by the Reshoring Initiative, the group that has spent 16 years arguing reshoring pays. The number that traveled came from page 7: 36% of OEMs have reshored or are actively reshoring, up from 29%. Manufacturing Dive ran both numbers.

On page 18, the survey asks the contract manufacturers, the job shops and fabricators who actually get the reshored parts, the same satisfaction question.

• 45% satisfied.

• 12% extremely

• 33% somewhat

The report’s own words:

“Reverses the assumption that contract manufacturers, who typically gain volume from reshoring, would be more satisfied than the OEMs bearing the cost and risk.”

So the shops receiving the work are less happy than the OEMs sending it, and the OEMs are less happy than they were a year ago.

The survey is small (249 respondents, 118 OEMs and 131 contract manufacturers, and the satisfaction question only goes to the ones who reshored, so call it 40 to 50 companies on each side). The contract manufacturer sample is about half fabricated metal. It is the same group that answered last year, asked the same question, and the answer moved 31 points. It is not a census. It is the first honest year-two number anyone has.

94% Of Lost Orders Are Lost On Price

The shops answer why themselves. When a contract manufacturer loses an order to an import, price is the reason 94% of the time (91% last year). Half of the shops that lost say the winning import price was 30% or more below their own quote. And they are competing against imports on more quotes than before: 38% of quotes in the last two years, up from 31%.

57% of contract manufacturers say the 2025 steel and aluminum tariffs hurt their ability to compete against imports. 15% say they helped. The tariff on the input landed on the domestic shop. The finished import did not pay it. Two shops in the survey said it plainly:

Majority of mold components are produced in China. Tariffs only increased our cost to make in-house and did not make us more competitive.

We have struggled to compete with Asian prices due to raw material costs. Our customers are paying prices for finished products that are lower than our raw material costs.

The first is an injection molder. The second makes metal components. Neither is named in the report.

Harry Moser, who founded the Reshoring Initiative and runs the survey, put the same math on the Industry Ignited podcast on Sept. 11 (episode 107). Say the Chinese part is $7 and the American part is $10. To win that order a US shop has to cut 30% out of its cost. If it is building a plant to do it, it has to cut the 30% and pay off the building at the same time. His words: “That’s pretty tough.”

The buyers in his own survey say the same thing. Asked how far US costs would have to fall before they moved 30% of their imports home under today’s tariffs, the most common answer was 21% to 50%. And 30 to 40% of them answered that they did not know. They cannot run the number at all. It is still hard to build a reshoring case right now.

Harry’s fix is to change what gets compared. Add up what the import really costs: the part, the ocean freight, the duty, the six weeks of inventory sitting on a boat, the rework when a lot comes in wrong. Put that against the domestic quote and the $10 part often wins. 40% of OEMs do that math now, up from 30%. The other 60% compare the domestic quote to an import price that does not include freight or duty, and pick the lower number.

The Work Came Back As Parts

What actually reshored, per the contract manufacturers: 69% of their reshoring cases were a customer switching component sourcing only, with assembly already in the US. 31% moved assembly and sourcing together. Same split as last year. A purchasing manager moved a part number from a Chinese supplier to a domestic one. The OEM quotes read like that:

Onshoring has moved manufacturing to a location with better cycle time which has improved our speed to market and on-time delivery. The additional work has increased our labor cost.

Difficult to find equivalent vendors, lesser quality, higher price, poor customer service.

By moving our procurement of glass doors from Taiwan to Kentucky or Texas we have been impacted mostly by labor shortages.

The positives are real and the report is right to count them:

• 70% of OEMs who reshored report faster speed to market,

• 65% better on-time delivery,

• 60% lower logistics cost.

The negatives are labor cost, vendor gaps, and hiring. 66% of all respondents rate hiring technicians (welders, machinists, electrical and chemical techs) as very difficult or a crisis. 60% say the same for maintenance techs.

And the shops that are winning say so, loudly, which is why the average hides them. A precision machining contract manufacturer:

We are receiving quoting activity that had been gone for 5-7 years and is coming back. Hit ratio is VERY high. Very rare that a requoted job is not won.

Jason Premo, who runs Acclaim Aerospace on Swiss lathes in Greenville, posted last Wednesday that about 90% of his RFQs over the last three months asked for rush delivery and offered to pay for it, mostly prototypes, across industries.

Russell Winter, who owns Center Tool Company in Illinois, published his insights on Sept. 4: 4140 bar for a repeat job up 40% in a year, price up 18%, order kept.

That is what “somewhat satisfied” looks like from inside a shop that is winning. The RFQ comes in, the material costs more, the price goes up less than the material, and the customer stays because importing the part takes six weeks. The same shop loses quotes on price and wins back jobs it lost in 2019.

Where a US shop does meet the import price, the survey and my feed agree on the mechanism. Jason Premo's LinkedIn headline says it in one line: small tight-tolerance parts, 1 to 32mm, “running 24x7 ‘lights out’ for China-busting low prices + aviation level quality.” On X, on a Sunday, it reads “the swiss lathes are machining parts lights out 24/7.”

That is one shop, one process, and a part size where a Swiss lathe with a bar feeder runs all night without a person. None of that is new. Bar feeders, pallet changers, probing and machine vision have been sold for decades.

The survey asked all 249 respondents whether AI and automation will make US manufacturing competitive enough to cut imports 50% by 2040. 33% said yes. 67% said no. The question puts two different things in one bucket. AI can read what a machine produces and learn from it. Loading the bar is still a bar feeder. The equipment that runs a shop lights out has been for sale for thirty years, and buying it is the part that did not happen.

A second survey this year points at what they are buying instead. CADDi, working with the Society of Manufacturing Engineers, asked over 200 US manufacturing professionals in January where their 2026 money was going.

• 69% said machinery and robots, up nine points from last year.

• 33% said operational software, ERP and the shop systems, down from 60%.

(CADDi sells software, so the second number runs against their own book.)

The same group put the skilled labor shortage at the top of the problem list, 79% of them.

So 69% are buying machines while 67% say automation will not cut imports in half by 2040. They are buying it anyway, mostly because they cannot find people to hire.

$135 Billion Left China

Kearney’s 2026 Reshoring Index came out in April. Direct imports from China fell $135 billion last year and China’s share of US manufacturing imports went below 10%, down from 20% four years ago. The other 13 Asian low-cost countries picked up $193 billion. The index stayed negative at minus 91. US manufacturing investment tripled over four years. Capacity grew 1.5%.

Rosemary Coates runs the Reshoring Institute, a consultancy that people mix up with Harry’s group. She said this on the PharmaSource podcast on Aug. 27:

If you look at the employment numbers, strictly employment in the manufacturing sector, it’s been absolutely flat for two years. So none of the political noise, the tariffs, nothing has really driven that big wave of reshoring that we were expecting and hoping for.

BLS has manufacturing up 58,000 since the December low and down about 34,000 over the year, which is her “flat.” Harry’s own count runs on announcements, around 250,000 jobs a year with 2026 tracking toward 338,000 based on the first quarter, and he said on the same podcast that hiring typically comes “a couple of years” after the announcement. Both numbers are right, because one counts press releases and one counts payroll. The survey sits between them: the announcements have turned into RFQs (32% of shops are quoting reshoring work, double last year), and the RFQs have not turned into a headcount or a margin yet.

Builder Read

If you own the shop. The work coming to you is a part number that moved, and a part number can move back. Half the shops that lost an order to an import lost it by 30% or more, and your material costs more than it did last year. Price the job for the day that import quote comes back in under you. And ask the buyer how they compared you. If they only looked at the price of the part, you are one quote away from losing the job.

If you buy the parts. Your supplier is less happy than you are. 45% of them say so, and they are absorbing material increases they cannot pass on. The ones who cannot keep absorbing it will reprice or stop quoting. Work out what your domestic part really costs you against the import once freight, duty and six weeks of inventory are in the number. 40% of your peers do that math now. The other 60% are comparing two numbers that do not measure the same thing.

I’m Watching This

• IMTS, this week in Chicago. Harry said on the podcast he hands out the National Metalworking Reshoring Award there. Every shop on my LinkedIn feed is on the floor at McCormick Place right now. The question to ask any of them at the booth: what share of your quotes go against an import, and by how much did you lose the last one.

• Oct 2, the September jobs report. Fabricated metal and machinery are the two lines that added people over the past year (BLS, up about 23,000 and 15,000). If the shops are at 45% satisfied because of price, the next thing to watch is whether they keep hiring into it.

Const

I’m Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I’ve seen every pitch this industry makes. Someone’s always selling hopium. I find it.

Friday's jobs report from the Labor Department added 16,000 manufacturing jobs for August. 12,000 of them came from the two lines I always read first: machinery and fabricated metal products. Machine shops, stamping shops, tool and die. The same table has the "compute factories" cutting 8,000.

Last week X spent five days arguing about whether a data center is a factory. I read those two lines instead. And they told me more than the whole argument did.

The Fight

Gavin Baker, who runs Atreides Management, wrote on Aug 30 that data centers “are actually reindustrializing parts of America and creating the kind of working-class jobs both parties have spent decades claiming to support.” Jensen Huang quoted him and went further: AI “is bringing manufacturing back to America.” The White House CTO called data centers “compute factories.” Danny Penny of a16z’s American Dynamism team wrote a long essay calling the data center “the forge” and named Hadrian as the proof, the same week a16z raised $1.1 billion for a Machine Age Fund.

On the other side, Mish Shedlock posted this chart:

Manufacturing construction down 32.6% in two years, data centers taking over, the revitalization “has failed.”

Under the White House’s jobs post on Friday

one reply said it shorter:

Still waiting for the resurgence to show up in manufacturing employment.

The skeptics have better numbers than the chart. The same jobs report that added 16,000 factory jobs has total construction employment down 93,000 over the year, data center boom included, because data halls are under 5% of private construction and housing is the rest. A $50 billion data center employs about a thousand people once it opens. And the machines inside it, the GPUs and the switches, are made in Taiwan.

The building is American. The tooling is imported.

$75 Billion Of Concrete

The concrete is going where Gavin says it is going. Census construction spending for July, out last week, has private data center construction running at a $75 billion annual rate, up 57% in a year. Private manufacturing construction is down 22% over the same year, and Mish’s number checks out. The Fed’s series is down about 33% from the peak.

Chip and electronics plants went from $97 billion a year of construction to $51 billion, because TSMC, Intel, Samsung and Micron finished the concrete on their fabs and moved to installing tools. Every other kind of manufacturing construction added together is about $117 billion, the same as a year ago. So what Mish’s chart shows is the chip plants finishing their buildings. Whether the policy worked or failed is not something I will argue here. But the 33% drop is chips, and only chips.

So the capex went to data halls, factory construction outside chips held flat, and the electricians followed the money: nonresidential specialty contractors added 8,000 in August. Gavin has that part right. What he does not have is a factory.

Metal Up, Everything Else Down

Here is the report, verbatim:

In August, employment in manufacturing continued its upward trend (+16,000) and is up by 58,000 since a recent low in December 2025. Employment in machinery manufacturing (+6,000) and in fabricated metal product manufacturing (+6,000) continued trending up in August.

And two paragraphs later:

Job losses occurred in computing infrastructure providers, data processing, web hosting, and related services (-8,000).

The “compute factories” cut 8,000 jobs in the month when metal shops added 12,000. That is one month, and one month is not enough to say anything. So I pulled the 12-month table, and that one is more interesting.

Over the past year, manufacturing as a whole lost 34,000 jobs. Inside that, durable goods added about 48,000 and nondurables (food, chemicals, plastics, paper) lost about 82,000. And inside durables, almost all of the growth came from three lines:

• fabricated metal up 23,000.

• machinery up 15,000.

• electrical equipment up 15,000.

• semiconductors, the CHIPS Act industry, lost 8,000.

• auto lost 4,000.

Those three lines are a shopping list. Electrical equipment is switchgear, transformers, busbars. Machinery includes the chillers and air handling that keep a data hall at temperature. Fabricated metal is racks, enclosures, brackets, the thousand machined and stamped parts inside a building full of servers. The jobs report does not say who the customer is. The shops do. All Metals Fabricating, a job shop in Texas, posted last week:

“We have seen a growing volume of copper busbar work in recent years, driven in large part by data center projects. To help keep pace with that customer demand, we have invested in a dedicated busbar production system that combines a punching machine and automatic loader with a CNC bending center.”

And on the MakingChips roundtable published last week (the data center part starts at 9:27), Mike Payne, who owns Hill Manufacturing in Texas, said January was the biggest bookings month in the shop’s 50-year history, February beat it, March beat that, and oil and gas stayed flat the whole time:

“Our ramp up has definitely been through the diversification, you know, some defense and outer space and data centers and stuff like that.”

Nick Goellner, who runs Hennig’s machine tool peripherals business, went further:

“That’s like the biggest reason for our growth. We’ve had this remarkable growth and it’s all because of data centers.”

The shop owners skipped the argument and quoted the busbars.

One thing does not fit. ADP, which runs payroll for about one in six private-sector workers, counted manufacturing down 17,000 in August while the government counted it up 16,000. Two surveys, opposite signs, same month. The government number will be revised twice more. August is not a resurgence yet, and it is not a collapse either.

One Hire In Every Tenth Shop

Small shops do not hire at the scale of a national number. So where did 5,700 fabricated metal jobs go?

Mostly nowhere in particular. There are roughly 55,000 fabricated metal establishments in the country (Census business counts) and 1.46 million people working in them. Adding 5,700 is four tenths of one percent. It is one hire in every tenth shop, or one shop in ten adding a second-shift operator.

The jobs report cannot split that between a big plant and a five-person tool and die shop, and I am not going to pretend it can. ADP’s size split says small firms under 20 people added jobs in August and firms of 20 to 49 shed them, across all industries, which is at least consistent with the smallest shops doing some of the hiring. That is as far as the data goes.

What the shops say on LinkedIn, where my feed is mostly small shop owners, is that the problem is not orders. Paul Van Metre posted last Friday that there are now waiting lists a machine shop can join for a chance to hire a graduate from a machining program, and that a shop he just recorded with could grow twice as fast if it could find people. Under it, Matt Guse of MRS Machining wrote that technical college programs are dying out and that the gap is less about skills than about the number of young people who want the work at all. So the work is there and the people are not. A shop that wants to double has one way to do it. One owner started posting his numbers last week, and his shop shows what that way is.

Six Spindles, One Shift

Russell Winter is a third-generation tool and die maker who owns Center Tool Company in northern Illinois. Last Tuesday he posted a photo of the floor, newest machine from 2019, and a promise to double revenue in public:

We are on target for $500k in ‘26 (pretty much been 500-800k since 1977, very unspectacular). In ‘27 we will break a MIL.

Two days later he published his scorecard. Four or five people including him. Gross margin last year 25% against a 40% target, net 6.8% against 10%. A repeat job that cost $386 in 4140 bar last year now costs about $540 in material, up 40%, and he raised the price 18% and kept the order. Six CNC machines, and his own utilization goal is written as six machines times five days, which is one shift. A reader counted the spindles in the photo and guessed he runs about 60% of that.

I have sold software into similar shops many times. Five people, older iron, one shift, material costs going up faster than his prices, and an owner who is the sales department. It is the typical American machine shop by headcount and by revenue, and there are thousands of them in the fabricated metal line. It does not appear on any investment tracker. The IndustrialSage tracker starts at $50 million per project.

One of the most useful replies he got was from Jason Premo, who runs Acclaim Aerospace on Swiss lathes lights-out:

Key will be unattended operations even if it is just nesting on vises or fixtures... anything you can do to get to 168 hrs/week of machining time.

Jason’s line answers the jobs question, and neither side of the X argument will like the answer. A shop at 60% of one shift does not get to $1 million by hiring five people. It gets there by running the six machines it already owns through the night with nobody there, and hiring one person who can set up a pallet pool and a probe. Russell doubling revenue adds one qualified job, maybe two. Do that in one shop out of ten and you get the small number the report printed, with more parts per person. Flat headcount, more output, and the new hire is a programmer.

Caleb Chamberlain of OSH Cut said the same thing from the software side last Friday, replying to a post about factories using AI to out-schedule each other. Scheduling, purchasing and production, he wrote, are best solved from first principles:

AI is the wrong tool in almost every conceivable way. AI can help write the software to do all of the above, though.

Hadrian built exactly that software in-house, more on that below. Russell’s second shift will run on a schedule, a fixture and a probe, with no model in the loop. The AI’s job in a machine shop is to help build that faster.

133 People At The Robot Factory

Danny’s essay leans on Hadrian.

The Arizona Commerce Authority release for the Mesa plant says $200 million invested and “will bring over 350 new jobs in early 2026.” Will bring. I could not find a current headcount for Mesa. For the whole company, LinkedIn shows about 600 people across four sites, up from 170 in 2024, and it raised $1.37 billion last month.

Anduril’s Arsenal-1 in Ohio is the other one everyone cites: 4,000 jobs promised by 2035. The first Fury drone rolled off the line on July 27 with 133 people on site. Anduril itself is a company of several thousand, but Arsenal-1 is the plant in Danny's argument, and the plant is 133 people 18 months after the announcement.

Both companies are doing hard, real work, and I am not arguing with either. My argument is with the claim that they explain the jobs number. Hadrian's whole company plus the Arsenal-1 floor is about 750 people. Fabricated metal shops added 5,700 in one month. And Danny’s line that “none of it runs without data centers” is his, not Hadrian’s. Hadrian describes its software as built in-house to turn CAD into machine instructions and schedule the floor. Nowhere does it say its plant depends on a hyperscale data hall.

So, Is It A Factory?

By the payroll count, no. A data center is a construction project that turns into a small utility. Meta’s Hyperion site in Louisiana is $50 billion for about 1,000 permanent jobs, with 5,000 construction workers at peak. Five construction jobs for every permanent one, and the permanent thousand run a building the size of a small town. Brookings looked at counties that got their first large data center and found on the order of 100 to 200 lasting jobs in a typical county, wages unchanged. Quincy, Washington, Gavin’s example, is the best case. Brookings measured what usually happens.

For a machine shop, a data center is a customer. It buys switchgear, transformers, chillers, racks and precision housings, and the three lines of manufacturing that make those things are the three lines that grew this year while the rest of manufacturing shrank.

So two things are pulling on the jobs number in opposite directions, and both are real. The AI buildout creates temporary construction work at a scale manufacturing has not seen since the fab pour, and then it creates permanent purchase orders for metal. The shops filling those orders survive on margins like Russell’s only by running unattended, so the orders turn into spindle hours faster than they turn into heads. More parts and better jobs, but not many more of them.

Builder Read

If you fund, supply, or hire into the buildout, the manufacturing jobs added in August were at machinery, electrical equipment and metal shops, not at a data hall and not at an automated defense plant with 133 people against 4,000 promised. That shop is paying 40% more for bar stock than a year ago and the one on record raised his price 18% and kept the order, so your tier-2 suppliers are repricing your BOM this quarter whether or not they told you. Manufacturing openings are at 580,000, up from 428,000 a year ago, which means that shop can decline your terms and still fill its spindles. And when a deck says the robot factory replaces that shop, ask for the current headcount, not the “will bring.”

I’m Watching This

• Center Tool Company, daily. Russell is posting statements, his scorecard, and what he calls the joys of defense contracting. A $500,000 shop going to $1 million in public is the most useful dataset on small-shop economics anyone has offered in years.

• Canada’s counter-tariffs. Canada now matches the US 50% tariff dollar for dollar on $27.6 billion of US goods: steel, dairy, paper, lumber. Russell’s 4140 is up 40% before any of that lands.

• Oct 2, the September jobs report. Three lines to read on one side, machinery, fabricated metal, electrical equipment. One line on the other, computing infrastructure providers. If they keep moving in opposite directions, that is the answer.

Const

I’m Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I’ve seen every pitch this industry makes. Someone’s always selling hopium. I find it.

American machine shops have spent forty years serving a short list of customers. Aerospace, defense, medical, and the big slow primes and OEMs that sit behind them. Everybody else, the small builder, the startup, the guy who needs eleven brackets, got a two week quote turnaround or no answer at all.

That was rational. It is also the reason there is no domestic supply base for anything else, and it is one of the holes China grew into. Not the only reason China won, of course. Wages, scale, state money, and Wall Street’s preference for outsourcing everything that did not move stock prices all did their part. But those we could not control from a shop floor. The quote desk we could.

A reader named Kev asked me to write about customer choice after I posted that picking the right customers is the highest-leverage decision a shop makes. So I put the question to operators. Many answered on the record, and the answer turned into a whole story.

The Short List

Look at which American shops are healthy. The ones on that short list. That is not an accident of taste. Those segments require AS9100, ITAR, ISO 13485, qualification cycles that run for years. A shop in Shenzhen cannot quote that work at any price. So the shops inside those segments can hold a margin, and everyone else competes with the whole world.

Caleb Chamberlain of OSH Cut wrote the mechanism in The Fabricator this spring, and put it in four lines on X in August:

It is a recipe for disaster. I call this the job shop doom loop.

1. Shops target large customers who do high volumes.

2. Large customers exercise leverage to squeeze margins and extend terms.

3. Job shops fill to capacity.

4. Job shops can’t grow because their cash is tied up in receivables and margins are slim.

Estimating, quoting, sales, AR and AP cost about the same whether the job is fifty dollars or fifty thousand. Spread that fixed cost and you need volume, volume means big repeat accounts, and big repeat accounts control your top line. Once they know it, the price goes down and the terms go out.

Now put the two together, because this is the part I do not think gets said. There was never a shortage of small customers. There was a shortage of small customers a shop could serve without losing money on the quote. So shops never saw that demand. The small buyer got no quote, gave up, or ordered from Shenzhen, and shop owners looked at their quiet inboxes and concluded these customers did not exist.

I spent twelve years attacking exactly this with DigiFabster, selling instant quoting to shops so the two week turnaround could become two minutes. Here is the honest lesson from those years: software alone does not fix it. A shop can install an instant quote widget and still run the same business underneath, same customer list, same terms, same fear of small orders. The quote got faster and the model did not change. The software only starts working when the shop changes around it: payment upfront, standardized work, an owner who actually wants the eleven-bracket order.

That full version is what SendCutSend and OSH Cut built. They went after the builder, the startup, the hobbyist, with instant pricing, payment upfront (OSH Cut runs about 80% card, 20% on terms) and parts in days. The demand was there the whole time. The tide is turning, and I want to be honest about how early it is: this is sheet metal, mostly, and most of what an American hardware builder needs at low volume still comes from Asia.

Why Sheet Metal Cracked First and CNC Did Not

OSH Cut’s framework is excellent and it comes from 2D. Flat stock, standard gauges, a bend table, a nest. The geometry is constrained by the process, which is exactly what lets software price it in a second and a machine run it unattended.

CNC does not hand you that. Neither does injection molding, and neither does an assembly with forty purchased components and a wire harness.

Protolabs is the cleanest evidence, because they publish the tradeoff. Their in-house CNC runs 3 days out of Minnesota and 4 out of New Hampshire, which is genuinely fast. Read the fine print and the New Hampshire envelope is 10 by 7 by 3.75 inches in soft metals and 10 by 7 by 1.75 in hard metals. Anything outside that box goes to their network, and the network runs 14 days standard.

You get the speed by constraining the geometry, not by wanting it. Sheet metal arrives pre-constrained. CNC has to be forced into a box before the same model works, and the box is small.

So when someone tells you the digital-first playbook fixes the industry, ask which process. OSH Cut now instant-quotes flat sheet, formed 3D sheet metal, tube, and as of recently bent tube. Every one of those is a process that constrains the geometry, which is what lets software price it. Simple 3-axis machining is partly cracked inside a size envelope. Complex five-axis, tight-tolerance, certified work is still quoted by a human, and that is where the short list lives.

One piece of the playbook travels everywhere though, and every shop should steal it: get paid up front where you can. Card, deposit, half of material and outsource cost before a chip is cut. It costs a customer conversation and it removes the entire problem the rest of this issue is about.

Buyers Still Price Like It’s 2010

I posted that most shops still run like the customer has nowhere else to go, and that customers do. Mike Payne, who owns shops and hosts Buy the Numbers, turned it around on me in the comments:

Fair statement. Reverse is true also. Lots of customers treat shops like they don’t have other options. They do.

Kirk Phelps, who hosts MachineShopGrowth, put a date on it in the same thread:

I think there’s a lot of buyers out there thinking they have the same weight to push around as they did in 2010, like people should be thankful just to have work, let alone make money. The supplier has more leverage right now than any other time through the last ~20 years I’ve been in this.

Worth noticing that “the customer” is two different people, and Caleb of OSH Cut staged the split as a dialogue:

Engineer: OSH Cut can put custom metal parts on my desk tomorrow.

Purchasing: Yes, but we need them to fill out a twenty page onboarding form, sign a custom NDA, and agree to our terms of service. Oh, and we want to hand-copy this quote into a PO, and pay them in 90 days.

Twelve years of watching quotes go in and out, I have almost never seen those two want the same thing. Purchasing still negotiates like it is 2010. Engineers do not. Fire a logo and you can lose an engineer who would have paid in fifteen days.

So what flipped the leverage at all? George Bain, who owns Bain Machine and Ronal Machine and bought both, answered it in that same comment thread:

The reality is Boomers are closing shops at record rate. I get no less than 5 Shops a month asking if I would like to buy them. If your a company who needs a great Shop to support your growth then there is less choice now.

George’s five a month is one inbox, not a survey. The published version is McKinsey’s: under current patterns, 92% of exiting owners simply close the business. I used that number in Who Owns Your Clock five weeks ago, and Mike Payne threw it back on X from the buyer’s side:

“That number should make every owner sick... buy the shop, install my people, my processes, my systems, and grow the value that was always sitting there.”

I have written about this wave a dozen times from the data side. This is the first time I have had an operator counting it from his own email. And notice George says closing, not selling. When a shop closes, the machines get auctioned and live on somewhere else, but the AS9100 approval, the customer qualifications and the trained crew have nowhere to go. They just end.

Net 90 Means the Shop Pays the Interest

I posted that net 90 is not payment terms, it is who owns your clock. Phil Timpson, a CNC applications engineer, replied:

net 90 is bullshit. I’m not a lending firm, here to give you an interest free loan.

Jonathan Sessa, who runs Fabworks, put the manners on it:

“Asking a small business for net 120 terms is the same as asking your local sandwhich shop for a discount. How do you not feel like a horrible person?”

It is worse than interest free, because the shop is usually the one paying the interest. The gap between shipping and getting paid comes off a line of credit, so the shop pays real money to let the customer hold cash for nothing. Average receivable balance times your LOC rate is the annual cost of being somebody’s bank. Run it before your next negotiation.

Caleb Chamberlain lived the other ending of the same story and posted it:

Made a concession and offered Net60 to a F100 company. We don’t usually do that, because it’s weird to finance the balance sheet of a company with billions of liquid cash on-hand. But they insisted, so what the hell.

Now 30+ days late on Net60 and emails to them disappear into the void.

Never again.

One shop refused the loan. One made the concession, to a customer with billions on hand, and still got paid late. “Get bigger customers and the terms get professional” does not survive either story.

Here is the war story. I asked Kirk Phelps whether he had ever actually fired an account negotiating like it was 2010:

100%. I walked from a ~$500k deal because of this exact thing. Cash flow is great, but when you’re talking about <10% margin on NET90 terms, you’re essentially paying your customer to make their parts.

Margin and terms compound. Thin margin on fast payment is a business. The same margin on ninety days means you financed the job yourself.

Notice what is actually happening in that story. The customer set the terms. The customer usually also sets the packaging, the labeling, the inspection paperwork, the portal you have to log into, and how often you may invoice. The buyer is not just buying parts, he is writing your operating model. Shops accept it because the alternative is an empty machine, and forty years later a lot of good operators are still running someone else’s business plan on their own balance sheet.

Ask them if it worked. I have been running an informal survey on X asking owners whether starting the shop was their best decision. The yes answers cite freedom. The no answers cite money. Not one has said the shop made them rich. Mike Bongiovanni, who was fired when the owner’s kids sold the company, started his own, bought that same shop back eighteen years later from the corporation that could not run it, and is twenty-six years in, ended his answer to me with

“Proud of what I’ve built but would I do it again??”

If you are about to buy or start a shop, that question is your #1 job. The grind does not fix a bad model. Pick the model first.

Say No, or Fund It

Say no means screening and it means having somewhere else to put the hours. Kirk Phelps screens and he is explicit that the screen is not price alone. He still takes thin work from customers who pay in thirty days or less, sometimes fifteen for the discount. The real filter is margin times terms times whether they actually pay. And the precondition: operating at 100% capacity is anathema to healthy growth, because the price you can charge and the lead time you can promise are not separate from how full the floor is. A full shop cannot fire anyone. Walking from $500k is a strategy only if you kept slack on purpose.

Fund it is the other half and the industry treats it as a dirty word. If your best segment pays in 90 days and drifts to 140, that is a working capital requirement, not a moral failure. It gets financed like any other: a bigger line of credit, an AR facility, a term loan, or outside capital. Price the carry into the quote, which is exactly what Payne does. He prices in terms, has lost jobs for it, and is fine with that, because “slow hurts more without question.” His bigger point, six acquisitions in:

Nobody goes broke buying machines. They grow broke funding the customer carrying material, labor, and regulation/compliance before anyone pays you a dime... A fortress balance sheet will be a key competitive advantage.

Phil Timpson gave the cheapest defense of the three when I asked what he runs instead:

net 30 is my go to, generally speaking. Signed quote with laid out terms, purchase order issued for records, signed service report, invoice.

Four documents before the invoice, each one closing a door a late payer could hide behind. It costs nothing but discipline.

Then there is the tax on all of it. Yasser Khan named it when writing about where small manufacturers could add capacity: they need help with certifications, paperwork and long payment cycles, because “a machine shop can’t hire ten people just to fill out forms”. Michael Kornuta put a number on the same thing last week, saying he meets job shop owners spending 20k a month or more on compliance and 1k or more on quoting software, and that half of them are drowning in work while the other half sit with open capacity. Caleb’s explanation of how you get to 20k was two or three hires, and he added that in a previous life doing government contracting the accounting department alone was a massive cost center just to hold DCAA compliance.

Enforcing terms is admin work. So is chasing money. A shop that cannot afford a front office takes whatever terms arrive in the PO, which is how the wrong customer walks in.

Prepaid is the escape hatch and it is not free either. Card upfront trades the customer’s float for the bank’s dispute desk: OSH Cut gets one or two chargebacks a year, the customer admits the error in writing, and they still lose the dispute. A smaller tax than net 90. Not zero.

Someone Is Living Inside the Spread

OSH Cut’s Caleb posted a comparison on Aug 29:

Those are his numbers on his own product, so read them as a vendor comparison.

Then someone in the replies asked how many people arbitrage the two. The answer was the part I cannot stop thinking about: Xometry names files a particular way, so he can see them arriving, and he knows a small company earning six figures of annual profit taking Xometry jobs and filling them with OSH Cut.

Somebody is making a living on the gap between what a channel charges and what a shop charges, and the only thing he adds is the click. That gap is quoting overhead, sales cost and payment terms, packaged and resold. It is the clearest price on the problem I have seen.

Jonathan of Fabworks has watched the same shift from inside his own order queue: a lot of these small jobs would either not have been ordered at all or gone offshore to China on longer lead times at similar cost. His frame is that the technology forces adaptation every decade or so, turret punches, then CO2 lasers, then fiber, and “software enabled high mix is likely the next.”

MakingChips clipped John Greeley of Navu saying quote response time is now the number one reason shops win or lose a job, because buyers go with whoever answers first. Slow quoting selects for the buyer willing to wait two weeks for a number, and that is usually the same buyer who sends net 90.

Kornuta asked the right question underneath it. Half his shops are overwhelmed and half are empty, inside the same forward-thinking group, so what is the differentiator? SEO, narrative, trade shows? Caleb’s answer was four words and a correction:

Customer mix, markets served. Manufacturing is technology, not a market.

Builder Read

If you buy parts, the 2010 posture is now expensive, and the certified segments you thought were captive are the ones with the shortest replacement bench. The shops with capacity, certifications and a working front office have the fewest reasons to take your terms. And early payment is the cheapest supplier discount on the market: I wrote up the collections data two weeks ago, where half of B2B invoices turn into cash in 25 days but getting to 90% takes 74, and the slowest tenth of $250k+ invoices takes 119. One shop owner offers 1% for net 10 and gets no takers, and asks for half the material and outsourcing costs upfront on big jobs because he cannot float anyone past 30 days. 1% for net 10 is roughly 18% annualized, and his customers walk away from it, because they want the float more than the discount.

If you run a shop, cut the quoting cost first. Every hour out of quoting widens the set of customers you can serve at a profit, and that is the only thing that makes saying no survivable. Keep slack on purpose. Get paid up front wherever the work allows it. And look at your top ten accounts as a receivables book, because that is what a buyer will value.

If you buy shops, George Bain of Bain Machine is describing your deal flow, and every shop that closes instead of selling is capacity nobody will ever buy. The terms on the top ten accounts will set the price long before the machines do.

I’m Watching This

1. George’s five shops a month. I am asking other shop buyers what their inbox looks like. If their rate matches his, shops are disappearing faster than any published number shows, and McKinsey's 92% projection is already out of date.

2. The $20k a month compliance bill. Two or three salaries of pure paperwork before a single chip is cut. If that is typical, reshoring is capped by the front office, not the floor, and new capacity will come online slower than the announcements say.

3. The Xometry arbitrage. Somebody clears six figures a year standing between a marketplace and the shop doing the work. Spreads like that do not survive: either the marketplaces reprice or the shops go take those customers directly.

4. CNC’s size envelope. Protolabs sells 3 day machining inside a 10 by 7 inch box and 14 days outside it. The whole instant-quote model is waiting on that box to grow. Whoever moves it takes the playbook from sheet metal into 3D.

Const

I'm Const. 12 years in US manufacturing: ran a shop, then spent a decade selling software to shop owners. I've seen every pitch this industry makes. Someone's always selling hopium. I find it.

For two months in this newsletter, I argued that America’s industrial middle is missing. This one is about who is actually filling it, who is pretending to, and what stays empty after all of them. I also asked the builders directly this week, in public, and their answers rearranged my map. And this newsletter is delivered to you today from 35,000 feet above the ground, so enjoy.

A Marketplace With No Machines Is Closing In on $900 Million

Xometry posted $229 million in Q2 revenue, up 41%, and raised its full-year outlook to 33 to 34% growth. Marketplace revenue grew 45%. Active buyers: 89,557. Xometry owns almost no machines. It is a routing layer between a CAD upload and thousands of shops, and buyers pay it close to a billion dollars a year mostly for one thing: not having to think about who makes the part (!).

And notice what that billion covers: mostly commodity processes, machining, sheet metal, molding, printing. The harder work, engineered components, assemblies, integrated builds, is not on the menu anywhere. A billion dollars for the "easy" slice (nothing in manufacturing is truly easy btw) tells you what the hard slices are worth.

Read that number the way an operator reads it. Xometry took years to build this, and it still grows 41% at this scale. Custom parts buying is nowhere near solved. The demand is real and the interface is the product.

The shops that actually make things keep proving the same point with machines on the floor. OSH Cut quotes around 12,000 parts a week with zero estimators, and cut lead time from 5 days to 2 in late 2024. The cut cost them 75% of rush-fee revenue. Growth roughly doubled anyway, and 2025 margin was an all-time high. OSH Cut's Caleb Chamberlain pegs the online-first fab market at $5 to 10 billion. And operators on X now argue the instant-quote model does not just take share from old shops, it creates orders that never existed before, work too small or too slow to quote the old way. So the pie grows.

Jim Belosic of SendCutSend says he is now building

“the largest, fastest, vertically integrated manufacturing company in North America... full-scale contract manufacturing that is accessible to EVERYONE”.

I love the energy. But everything that made this tier work so far came from doing less, not more.

So execution is getting solved. Upload a file, get a price, get the part. That was my decade (after exiting my shop, I spent twelve years selling quoting software to shop owners), and watching it become normal is strange and satisfying at once.

Also, feels like the execution was never the whole problem.

The Moonshots Keep Publishing the Shopping List

Look at what the ambitious American hardware companies build in-house. Hadrian raised $1.37 billion at a reported $8 billion valuation to industrialize precision machining inside its own walls. Musk’s Terafab is a reported $16.8 billion first phase to make chips the way SpaceX builds rockets, everything under one roof. Tesla designs Optimus actuators itself. Same with every part Anduril pulls in-house.

Each of those budgets is money the founder would rather have paid to a supplier. But the supplier does not exist. So the budgets became the shopping list.

The bull case for this model has a name. Packy McCormick’s Vertical Integrators series argues the winners of the techno-industrial era will own their whole stack:

“Aggregators won by owning demand; Vertical Integrators win by owning supply.”

His thesis in one line: there will be a SpaceX and an Anduril in every big category, and they will be bigger than the incumbents they replace. For any single company, he is right. When your supplier ships you a faulty thruster, you build the thruster. Integration is the rational answer to a base that cannot supply you.

Now let’s zoom out one level. Shenzhen solves the same problem with a catalog. A founder there designs from parts that already exist: the reducer, the encoder, the gripper module, each with a datasheet, a price and this week’s lead time. COTS from a catalog, custom from the shops nearby, and you assemble a product like industrial legos. America has the COTS half (shout out to beloved McMaster-Carr) and is building the custom-execution half (the Xometry and OSH Cut tier). While the middle of the set is empty: precision electromechanical components, reducers, precision motors, encoders, the parts whose moat is process knowledge rather than patents. In the visible funding data, the megarounds are all machines and brains, almost never the parts inside them.

My test for whether this tier exists is simple. Can a three-person team order the component today, CAD file behind a download button, no MOQ, no NDA, shipping this week. Today the honest answer is usually Alibaba. American robotics startups fly components home from China in suitcases, and some buy Unitree humanoids just to strip them for parts (a claim Alvin Graylin repeated, as rumor, on the Moonshots pod last week: chips get smuggled east, actuators get smuggled west).

I Asked the Builders What Is Missing. No Two Answers Matched.

This weekend I posted one question on X: what is the one missing piece of the US ecosystem that stops you from building end products locally?

The first answer came from Resolved Motion, a young company making NDAA-compliant robotic actuators in upstate New York, which is exactly the tier from the section above. Their answer, in full: “Capital.” I asked what kind, VC money, a bank loan, or customers paying upfront. Their reply:

“None of the described asset classes are good for that pre-revenue. VC has forgotten what they do, banks need collateral, customers need inventory to buy. We need new asset classes.”

That answer corrects something I was about to write. On Aug 20 the SBA proposed calling a machine shop small at up to 1,900 employees, up from 500, which hands the 500-to-1,900 tier set-aside access and SBA 7(a)/504 debt at the same moment. I read it as the first financial instrument shaped for the component tier, maybe by accident. Resolved Motion’s point stands anyway: being eligible for a loan and getting one underwritten pre-revenue are two different problems. The machines come before the revenue. The asset class for that does not exist yet.

The other answers, each from a different direction. My good friend Matthew Chang:

“World class automation/manufacturing engineers.”

The people who can design this tier are at Moog, at Nidec, or retired.

Anthony Sardain:

“Factories that answer their emails.”

Five words that explain most of Xometry’s revenue.

LightGage:

“Would be beneficial to have a highly predictable tort system.”

I pushed on this one and the answer got darker:

“It’s an asymmetric bet across all liabilities (e.g. product, IP, labor, regulation, taxing authorities, etc) that you make essentially unknowingly that can lead to complete bust/ruin... Insurance companies use exclusions to avoid risks with unknown or unmeasurable probabilities.”

An unmeasurable risk does not get priced by insurers, it gets excluded, so the owner carries it themselves. Their sharpest point: the damage is the businesses never started, because nobody can compute the depth of the downside.

One more reader added tools to the list (robots cheap enough that a small maker competes with a factory).

Where This Is Actually Heading

Put the pieces on one table. America now has demand routers (Xometry), disciplined productized makers (OSH Cut, SendCutSend, Fabworks), venture fortresses (Hadrian, Anduril, Terafab), operator roll-ups quietly buying retiring shops to build (or ruin?) mid-size players (a path I will cover separately), and the first merchant component makers, which are born into a capital vacuum. And five different missing layers underneath all of them.

Destin Sandlin’s new SmarterEveryDay episode adds the missing mechanism from the other end: he tried to source a simple all-American product and found the gap is not engineering, it is decades of margin optimization hollowing out suppliers one purchasing decision at a time (his word for the end state of that optimization is “encrappification”). The same math that hollowed the middle will not rebuild it. Rebuilding needs actors playing long-term economics: automation, volume over margin-per-unit, catalogs over quotes. The digital-first shops already run exactly that playbook. They are the proof it works at American labor costs.

So here is my honest picture of the end state, and it is neither Shenzhen nor the fortress. It is groups of narrow, productized, software-front specialists that plug into each other: parts companies feeding assembly companies, catalogs a founder can design from, manufacturability checked by software at upload before a bad design ever reaches a machine. AI design tools are already racing toward that last layer (Autodesk ships AI DFM reports in Fusion, startups are putting manufacturability inside CAD), and whoever encodes real manufacturability decides what AI is allowed to design.

Industrial legos as an ecosystem of specialists and not one company that ate them. The missing pieces are the five the builders named: an asset class for pre-revenue machines, the engineers, a computable liability system, interfaces that answer, and cheap enough automation.

One thing my issue does not argue: that imports are the enemy. America needs to be able to do much more at home, and keep cooperating with the world at the same time. I strongly believe normal trade with China will continue in any future I can imagine. The gap that matters is narrower: whole categories of components where no domestic option exists at any price, at any lead time, from anyone. The goal is having the option, not making everything ourselves.

I’m watching this.

• China displaced Germany as the world’s largest machine-tool exporter in 2025, EUR 8.6 billion vs EUR 7 billion, first time ever. The stuff that makes the stuff now ships from China.

• Shanghai’s STAR market listed Unitree at a $9 billion pricing (closed up 460%) days after LandSpace landed an orbital-class booster, punching the ticket on an IPO lane that requires a landing, not revenue.

• The new National Security S&T Strategy names “strategic capital investment structures” for commercial technology in national security. Fourteen priority technologies, zero dollars attached so far.

If you operate in any of these five layers, hit reply or DM me. I read everything, and the best answers already changed this issue once.

Const

Jim Cantrell was employee number four at SpaceX. Now he builds rockets at Phantom Space, and last week on the Powers podcast the host asked him a simple question: why build your own rockets at all?

His answer:

“There’s not a great supply chain out there. These tanks sitting behind you here, we could find somebody to build them, but is that their priority in life? No.”

And then he added:

“There’s a layer at which the supply network is adequate, and then there’s a layer at which it’s not. And that’s why you wanna vertically integrate, is where that supply chain is not adequate.”

These days vertical integration is the most celebrated feature of the American buildout. But in Cantrell's version it is a workaround. You integrate where the suppliers don't exist. In Shenzhen a company integrates by choice, sitting inside the densest supply base on earth. Here, in the US, a company integrates because there is literally nobody to call.

This is the missing middle I keep writing about: companies that own their component design (a reducer, an actuator, a valve), keep it in stock, publish a catalog, and make the part better with every customer. US robotics startups raised $18.8 billion so far this year. Every round above $200M is a robot, a vehicle, or a foundation model. Component makers got exactly one round worth mentioning: PaXini, tactile sensors, $145M. PaXini is in Shenzhen. America's counterplay, Westmag, raised $11 million and said it themselves:

“Right now America’s share of that muscle is essentially zero.”

Too Much Money Kills Companies

Cantrell spent three decades watching rocket companies die, and he keeps the score simple:

“If you look at the history of rocket companies, the ones that have done it have actually done it for a hundred to two hundred to three hundred million dollars max. Firefly, SpaceX, Rocket Lab, they’ve all been super cheap. The ones that did it for billions, Richard Branson, he’s gone out of business. Too expensive. ABL, they raised almost a billion dollars. They failed. Astra raised several billion dollars. On and on and on. Everybody that’s raised that kind of money had people expecting them to spend it.”

“Had people expecting them to spend it” — a very striking way to say it loudly.

His failure numbers are from memory, so treat them as his numbers, but the pattern is real. And he thinks the money itself was part of the problem:

"You end up raising a lot of money, but it's the wrong kind of money."

Also, Phantom doesn’t announce funding rounds anymore. I strongly believe this is a good reason why not to do that.

“The army of looters that show up at my doorstep take months to sort out. So what good is it announcing a funding round? We make the filings, and people eventually figure it out.”

The mega-round announcement is the main marketing instrument of this entire wave. Defense, drones, humanoids, all raising billions on reindustrialization slogans. I think Cantrell's rule holds here too, which means most of that money buys burn rate, not capability, and this wave produces the same body count the rocket wave did. The survivors will look like survivors always look: lean, grinding for a decade, profitable before famous, raising later on their own terms. This is not a doom take (the rockets got built, space got cheap). But the money and the capability are different things, and this money is buying final-product logos while the tier underneath stays empty.

Integration Is a Symptom, Not a Cure

Justin Lopas ran production at SpaceX and Anduril and now co-founds Base Power. On Powers a week earlier he gave the make-versus-buy logic: build in-house where you add value a supplier can't, where you won't outsource quality, and where you need what he calls security of supply. Base leans "more towards the Tesla side" because vertical integration "reduces cost meaningfully, and cost is the major driver of our economics."

All of this makes sense for his current business. But notice what happens: no new suppliers appear. A vertically integrated plant answers its owner's questions and nobody else's. Terafab, the $16.8B building in Grimes County that people calls America's Shenzhen, is one owner under one roof by design. Actual Shenzhen is the opposite: Seeed Studio's Eric Pan calls it "a jungle with enough biodiversities", a base dense enough that a four-person team walks into a factory and starts from what it already builds. His method:

“It’s not ‘design for manufacturer’, it’s ‘design from manufacturer’.”

You cannot design from a supplier who does not design. A megafab does not create a thousand suppliers, it makes them unnecessary.

And before someone brings up BYD (batteries, chips, even the seats): BYD integrated from inside the densest component base on earth. That is how its Hungarian plant hits full output this quarter, about 18 months after ground broke. Our integrators are assembling their verticals from imported machines and PDF catalogs. Integration is downstream of density, not a substitute for it.

Busy Shops, Empty Middle

Every shop owner reading this is about to object: we're busy. I know, I hear the same from every operator I talk to, and the data agrees. Core capital goods orders were up 12.5% year over year in June, the highest since November 2021. AMT economist Christopher Chidzik, on the Quality podcast in June:

“When the USMTO orders go up, that’s when people want to make things. When the workholding and tooling are going up, that’s when people are making things.”

Both are up right now.

Jim Belosic of SendCutSend explained where the demand comes from, on the Relentless podcast in May:

“People are a little scared of offshore now. Are they gonna cut off my supply chain without any warning?... The US is actually a viable option now. We’re in our early days for accessibility and price and speed and everything, but we’re getting there. People are testing the waters.”

But capacity, in my view, was never the problem. The US has plenty of machining capacity for its own demand, and reshoring just adds more work on top. None of that work changes what the shop is. A shop booked solid cutting someone else's prints is a busy print-taker (my corner of the industry, and I say this with love). Cantrell again, in the same answer about tanks:

“Do we machine these valves? No, I don’t wanna be in the machine shop business.”

The shop gets the valve order, and the design stays inside the vertical company. Busy and dense are two different measurements.

So what happens to the machine shop? It survives. The comfortable version of it (wait for prints, quote, cut, ship) gets squeezed from both sides, integrators pulling strategic parts in-house and automated shops compressing the commodity work. The way out is to go vertical in the small: own one part. A gearbox, a spindle, a fixture line, bought from a retiring competitor or grown out of your own niche. And one honest warning before anyone runs with this: nobody here will beat a Chinese catalog component on price in the open market, so the part you own has to live where the Chinese version can't follow you. Defense and other critical industries that need a domestic source and pay real money for it, or customers who pay for a one-week lead time instead of a six-week boat.

This lane is narrow, and owning a part takes much more than machining capacity (the design, the testing, the inventory, the catalog, the support). But that's a densifier at shop scale, and that's the version of vertical integration that adds to the supply base instead of eating it. And I know some brave ones who are building towards this direction (kudos!).

The Ten-Year Question

So how does this look in ten years? I see three endings for the component tier.

One: everyone builds in-house forever. Innovation happens behind each company’s wall, priced for one user, and dies with the program. The defense supply chain already works this way. It works, for customers who can bill a government.

Two: we keep buying the middle from China and feel fine about it. Froman laid out the problem in Foreign Affairs last week: Chinese producer prices negative for 38 straight months, roughly 30% of industrial firms operating at a loss, and the state keeps them building. Solar cells: exports up 73% in the first half of 2025, unit price down about 25%. A supplier whose prices fall every month feels like a gift. Then come the export controls, or the breaking point Froman predicts. Either way, the decision gets made in Beijing or Washington, not by any of us.

Three: the middle actually forms here in the US. Components re-engineered (reverse IP acquisition) from Chinese and Japanese origins into American catalogs, funded the way Cantrell describes (order books and debt, not billion-dollar equity), behind whatever tariff umbrella survives this decade. Slow and boring. Also the only version where a small team in Ohio can build a robot from parts it can actually buy.

How fast we get to any of these depends on China… Either it grinds on, imported components get cheaper every year, and American component makers survive only where tariffs protect them. Or it breaks, and the crisis takes everyone’s capex down with it, because nobody funds new component factories in a crash. Neither ending hands us the middle. The middle gets built on purpose, or it doesn’t exist.

Builder Read

If Cantrell's rule works for this wave too, the best bet is not the $2B humanoid round. It's the $10-30M company with real orders, its own catalog part, and no investor pushing it to spend faster.

Watch what the integrators decide to buy versus build: every part that Base Power or Anduril buys from outside means a component market just opened, and every part they pull in-house means one just closed. The concentrators are already expensive (some trade at 44x an estimated revenue). The densifiers have no price yet because nobody funds them. Same story as last week: do the humanoid companies buy actuators or build them? And if you run a shop, the answer is not more capacity. Find one part you can own.

I’m Watching This

1. SpaceX rideshare. Reporting says it stopped taking Transporter and Bandwagon bookings past late 2028, sourced to satellite operators, with no SpaceX statement. Cantrell said the same on Powers. If it holds, every smallsat builder loses its bus ride, and the launch middle reopens overnight.

2. The AMT pairing. Machinery orders just had their strongest half since 1998 ($3.44B in H1). If they flatten from here while tooling keeps climbing, shops chose sweating existing machines over buying new ones, and the capex wave stalls at the equipment tier.

3. The G-20 in Miami this December. Bessent already said the world "cannot take a China with a trillion-dollar trade surplus." If China's surplus makes the official agenda, the pressure on Beijing becomes real policy, and just keep buying from China stops being a stable plan.

Const

Somewhere this week an engineer from one of the OEMs needed five flat parts bent and coated, and another one needed a shop that holds two tenths on titanium. Even ten years ago that search started with a buddy he knew for some time or a trade show booth, or a bookmarked directory in the browser. Now it starts with a question typed into an LLM. The engineer reads one composed answer, checks two links, emails (or uploads a part for instant quote if he is lucky enough) three shops. If the LLM has never scraped anything about you from the Internet, you were not one of the three.

Manufacturers walked into this one. The industry skipped the SEO era (and arguably the whole Internet thingy) almost completely. Thirty years of word of mouth, trade shows, and a website somebody’s son made once. David Petrich of Machinable Innovations described his whole marketing stack with total honesty on Machine Tool Maniacs in June:

Word of mouth, a decent website, google reviews so an unfamiliar buyer thinks “this place is legit,” and social media, “terrible at it.”

— his words.

I sold software to shops for twelve years. That description covers much of my old customer base. And it worked, because the buyer was a human who lived nearby and asked other humans.

The modern buyer is not that anymore.

The Long Tail Never Touched Manufacturers

Randy Altschuler, Xometry’s Founder (Xometry owns Thomasnet), said the quiet part on Machine Shop Mastery last month:

The way people search is changing with LLMs, and “that long tail of the internet had really not touched manufacturers.”

The internet found restaurants, retail, your next vacation spot. It never really found the 50-person shop, because the shop never wrote anything down in public. His advice: digitize what you offer, put it on the site, keep the parts catalog alive.

Notice what a live catalog is. It is the thing a model can cite. A capability written in plain language is a fact a model can repeat. A brochure PDF from 2014 is neither.

The mechanics are unforgiving and worth understanding. An LLM does not rank ten blue links and let you fight for position three. It composes one answer from a handful of sources it considers credible, cross-checks claims between them, and hands the buyer a conclusion. No clicks, so your analytics show nothing. Nothing shows up as a lost quote, because the RFQ never reached you.

The discipline of fixing this has a name now, AEO (Answer Engine Optimization): showing up in the answer instead of the rankings. Matthew Bernard wrote the sharpest thing I have read on it this week in Industrial Base Alpha, and this issue leans on his framing throughout (thanks much for inspiration, Matt!).

The Founder Is the Marketing Department Now

Here is the part our industry likely will resist the most. The strongest signal a model finds is a named human explaining the work in public, repeatedly. Not a brand account posting spec sheets. But a person.

Look at who already runs this play. Chris Power narrates Hadrian on every podcast that will have him, and the transcripts of those podcasts are now training data. Palmer Luckey’s interviews get written up by five outlets each, and every writeup is a citation. In our corner of the industry, Jim Belosic turned SendCutSend’s shop floor into daily content years ago, and OSHCut’s Caleb Chamberlain writes like an engineer talking to engineers. When a model composes an answer about laser cutting or defense machining, these are the voices it has read. That is not an accident of charisma (but there is a component of that ofc). It is volume of public text.

The capital side already treats this as infrastructure and not a vanity. Harry Stebbings is telling founders that not posting daily is “a massive mistake”.

Chamath just published 108 pages arguing creators are the new demand generation for “trillions of dollars of goods and services.” And Pieter Levels reports spotting people worth $500M and up buying X followers, half a million a month, engagement flat, completely fake.

Audience became a currency the rich buy directly. But you cannot buy your way into the answer, because a model never sees the follower count, only the writing under it.

Defense is just the extreme case, and it is where Matt’s piece goes deepest. His observation is that

“The smartest people in Washington are turning to Large Language Models to bring themselves up to speed,”

and the reason is arithmetic: the Golden Dome IDIQ approved 2,440 vendors out of 2,463 applicants. Nobody tracks 2,440 companies by relationship, so the staffer shortlisting five names asks a chatbot, and Congress already spends 88% of its AI tool dollars on ChatGPT (CNBC, via Bernard). His finding is that Anduril shows up in those answers more than Lockheed, because the founder talks in public and the model reads. His warning to the primes:

“If your entire government affairs budget is going toward people and not toward media, you are playing catch-up. You are letting an algorithm tell your story before you even enter the room.”

But read his piece from a job shop’s POV and it gets more interesting, because you never had lobbyists. For the rest of manufacturing this was never about influence. You had no lobbyists to replace. It is about being findable, and the numbers are lopsided: a good conversation at a booth reaches one buyer, while a capability page written once gets read by the thing every buyer now consults before calling anybody.

The Answer Engines Do Not Read the Same Internet As You

One more layer, and it changes the plan. The models are not reading one shared web anymore.

• X locked its data years ago, blocks the crawlers from OpenAI and Anthropic, and feeds everything to its own model; since the xAI merger, Grok is the only model that reads X in full.

• Reddit licenses its archive to Google and OpenAI instead of giving it away.

• LinkedIn sits inside Microsoft, which decides who else gets to read it (why there is so much AI slop in it now).

Live example from this week: the best defense market commentary I read was a 15-tweet state-of-the-market thread from Nathan Mintz, 21 years in old and new defense. It lives on X. Exactly one model will ever learn from it.

For a manufacturer planning a public presence, that has a practical meaning. Post only on X and you are training exactly one model, the one your defense-tech buyer might use, and none of the others. Post only on LinkedIn and you are betting on Microsoft’s stack. The one corpus every model still reads is the open web: your own site, a blog, a Substack, podcast transcripts, trade press.

So write it once on a domain you own, then let the platforms carry copies. Whichever model your buyer asks, at least one route to your writing stays open.

Anyone Can Feed the Answer, Including a State

This week Zane at Nox posted a screenshot with one word, "Interesting." Under it, an LLM answer comparing American and Chinese tax burdens: confident numbers, clean bullets, and one citation: a China market-entry consultancy's blog.

The numbers were right. I checked. OECD puts China at 20.4% of GDP for 2023 and the United States at 25.6%, and the screenshot was within a rounding of both. The consultancy had done nothing wrong. It read the OECD, summarised it correctly, and got named for it.

Then I tried to open the OECD reports myself and could not. Both are PDFs sitting on a path that blocks automated fetching. I downloaded them by hand.

So the better source is accurate and invisible to the machine doing the citing. The consultancy’s page is plain HTML any crawler can read. The consultancy is in the answer, and its brand rides along every time that screenshot gets forwarded as evidence. The thing that decided who got credited was not accuracy but whether a crawler could open the file.

Shops lose citations the same way. Your good information sits in a PDF brochure, a scanned AS9100 certificate, a downloadable line card, tolerances that appear only after a quote form. A model cannot read any of it. So the competitor who typed the same facts onto an ordinary page gets cited instead, whether or not his shop is better.

If a consultancy can win the slot, so can a state. NewsGuard audited ten leading chatbots and found they repeated false narratives from the Pravda network 33% of the time, with seven citing its sites as legitimate sources; that network published 3.6 million articles in 2024 across 150 sites (Forbes). The technique has a name, LLM grooming. Nothing in that pipeline is disclosed, so you cannot tell who fed the answer in front of you, and what to do about that is a policy question I won’t litigate here. For a manufacturer the practical read is narrower: the answer layer is contested ground, and it rewards whoever leaves the most consistent public record.

There is one good thing in all of this. A model wants to see a claim in a few places before it repeats it, so anything you said once, on your own site only, tends not to stick. Saying true things in more than one place is both the decent way to do it and the one that works.

Builder Read

Go run the test tonight, it costs you nothing: ask three different models who should make your kind of part in America, or who leads your category. If you are not in the answer, that is the pipeline you are not in, and no sales hire fixes it.

Then look at what Jon at Fabworks has been doing, because he is already many years ahead of most shops and probably did not set out to do it for the machines. Design guides on K-factor and kerf and distortion zones. Material comparisons, gauge charts, tap and drill charts, export instructions for Onshape and Fusion. Tolerances and lead times stated in the open instead of hidden behind a quote request. Add the engineering videos he posts on X and LinkedIn and you have a company that has written down, in public, most of what it knows. Every one of those pages is a fact a model can repeat about sheet metal, and his name is attached to all of it. Be like Jon.

The rest is boring and it compounds. Capability pages in plain language that answer the questions buyers actually type. Specs, tolerances, materials and the catalog live on the site, not locked in brochures. The same company facts everywhere the company appears. Get on the podcasts and into the trade press, because transcripts are what models read. And the founder writes, on a cadence, as a person. A booth works four days a year, but your answer works every night. Assign one person to own what the machines say about you, the way you assign someone to own quality. Your competitor already did.

One important limit on all of this, from the same Mintz thread: the buyer still concentrates money in whoever can actually deliver. By his account, one Pentagon task force studied counter-drone for a year and a half, then put roughly 60% of its budget into two established names, and Lockheed’s latest PAC-3 award was “larger than all OTAs in last 2 years, put together.” His summary of what the customer buys:

“Industrial capability, not novel tech. The tech is good enough.”

So keep the two jobs straight. Visibility gets you into the shortlist. The factory gets you the award. AEO is the front of the funnel, not the funnel.

I’m Watching This

1. Matt promised AEO visibility benchmarks for the drone industry within a week. First dataset of its kind; the methodology matters more than the rankings.

2. Xometry rebuilding Thomasnet “for the searchers.” If a 30-year directory becomes the layer LLMs cite for supplier discovery, that is a position nobody priced.

3. The follower market. Levels’ observation says audience is now a store of value for people who already have everything else. Watch what that does to the value of any follower count, including yours.

Const

I made that call in May: China spent 30 years and 700 million people assembling that machine, and no subsidy rebuilds it on a ten-year horizon. Now I can put prices on it.

In 2002, Chinese factory labor cost 57 cents an hour. A hundred million people were moving from villages into factory cities. The whole world was opening its wallet. Those three conditions built Shenzhen’s supply density, and none of them exist anymore, anywhere. Not in Vietnam. Not in Mexico. Not even in China, which is why BYD’s newest plants are in Thailand, Brazil and Hungary.

So the low-wage path is closed. Not hard. Closed. No amount of subsidy reopens it. The conditions are gone.

But there is a second path, and three countries have been walking it this whole time. In 2012, the last year the BLS ran the comparison, German factory labor cost $45.79 an hour against America’s $35.67, with Switzerland at $57.79. They pay their factory workers more than America does, and they are still manufacturing powers. Japan pays about what we do and runs the deepest machine-tool bench on earth. How? None of them do consumer assembly. They make machine tools, precision components, instruments, factory automation. The stuff that makes the stuff. High knowledge, low labor share, number one in a thousand narrow markets that never make headlines. The Germans even have a name for those companies: hidden champions.

That is the whole choice. The low-wage path is gone, and the high-wage path is not a consolation prize. It is the layer with the margins, and it is the national version of what I keep saying about shops in Stop Counting Heads: fewer people, much more per person. America already produces $220k+ of value per manufacturing worker against China’s $40-50k. We are already the high-wage producer. The mistake is spending policy and capital trying to become the low-wage one again.

And America already owns pieces of this layer:

• Applied Materials, Lam Research and KLA build the machines that make chips; most of the world’s fabs stop without them.

• Lincoln Electric has made welding equipment in Euclid, Ohio for over a century and pays factory workers a profit-share bonus big enough to make national news: $33,029 a head for 2013, the last year they published the figure (hope it’s still going well).

• Hypertherm cuts steel around the world and is owned by its employees.

And the next generation is already forming:

• Machina Labs teaching robots to form sheet metal in Los Angeles.

• Divergent printing and robot-assembling structural parts, and raising $290 million in September 2025 at a $2.3 billion valuation to scale it for US defense.

• Agility building robots in an Oregon factory built for building robots.

Three stories from this week show where the line runs.

America Is China’s Third-Biggest Customer

China’s exports to the US fell 20 percent last year, and total Chinese exports still grew 6.1 percent.

Where the goods go now: China’s total trade with ASEAN was $1.06 trillion, the EU $830 billion, the US $560 billion. America is third.

Some of that is relabeling, the Mexico story from May at national scale. But relabeling is the small part. Global Trade Alert sized the redirection at about $150 billion, landing in ASEAN, Africa, Latin America and the Gulf, with none of it showing up in the EU or Japan. Those are not transshipment hubs for the American shelf. They are customers.

The companies read the same way. Shein’s Europe business is now bigger than its US business, $14.8 billion against $10.1 billion, and the US side has sat flat at about $10 billion for three years. BYD sold 1.04 million cars overseas across 110 countries without selling one passenger car in America.

A tariff wall is a bet that reaching the American consumer is the thing worth paying for. That bet gets cheaper to lose every year.

The Wall Went Up Before the Supply

On Tuesday the FCC put foreign-made robots and power inverters on its covered list. The exemption path runs through the Buy American standard, and that standard has a schedule: domestic components have to clear 65% of component cost through 2028, then 75% from 2029. Assembly does not count.

Three weeks ago, in What Survives to 2036, I put humanoid robotics at 40 to 70 percent Chinese BOM. This week the FCC made that number a compliance problem.

A robotics builder did the math in public: you cannot buy American actuators at scale. The one part every robot needs. Thirteen hours into his thread, a small shop in Upstate New York replied that they build precision rotary and linear actuators for humanoid robots, in their own plant.

Hold both facts. There is a legal requirement, and there is no supply. Whoever builds the supply has customers waiting by law, and the deadline is 2029.

The timing matters more than the rule. Automation is the only way a country with American living costs competes on labor, and the US just restricted the imports its own automation runs on. Robots for robots come first. And the honest benchmark for a robot here was never “beat a Chinese worker.” It is “beat landed cost plus tariff plus six weeks of lead time.” That is a lower bar. In some categories it is beatable now.

Notice what an actuator is, by the way. It is not a consumer product. It is the stuff that makes the stuff.

The Machinist Isn’t Underpaid

A shop owner telling kids to stay out of the trade he runs. Downthread he says why:

“the wages are still stuck in the 1990’s.”

The two ads are not a clean comparison, the Domino’s number folds in tips and mileage against the driver’s own car. The national medians are cleaner and they say the same thing. Machinist $56,150, plumber $62,970, same high school diploma to start, and machining is the only one of the two projected to shrink. Meanwhile 462,000 manufacturing jobs sit unfilled.

Everyone reads this as “manufacturing pays too little.” I read it the other way.

The machinist is not underpaid by world standards. He is priced out by his own economy.

Housing, healthcare and school set the cost of his life, and the services economy around him sets those prices. An expensive life prices labor out of factories. A tariff adding 10 to 12.5 percent to a bike, depending on where it comes from, is real money, and it is small next to rent.

The German machinist costs even more than ours and stays busy, because he works in the layer where labor is a small share of the product and the product is number one in its niche. That is a positioning problem, not a wage problem. Fixing the machinist’s rent is government work, not startup work. Fixing what he works on is ours: fewer people, much more per person, in products the world has to buy.

A Wall With No Export Test

One warning before my Builder Read. Japan's phone industry grew up inside a protected home market and never won outside it. Korea protected its industry too, with one difference: survival was conditional on selling abroad. POSCO and Samsung had to win outside the wall to keep their place inside it.

The sharpest blueprint for rebuilding here makes the same miss. Aaron Slodov’s “How to Build American Shenzhen, Fast” is the best case anyone has written for rebuilding the full stack, and I agree with most of it (his factories run on software and small teams, his list runs materials, electronics, machining, actuators). But read it asking one question: who buys? The answer is American startups and the Pentagon. The word export appears four times in the whole essay, every one about Chinese export controls. His stated endgame is that Chinese component dependency “becomes moot.” That is import substitution, not competitiveness.

And his map of 68 reindustrialization companies has exactly one entry under machine tools, a mobile machine shop in a shipping container. He sees the gap. He writes that America "barely makes the machines that its factories run on," and then scores that category a 3 out of 5 on venture return and a 3 on dual-use, in the same entry where he says the dual-use case is inherent because every factory needs machine tools. It lands twelfth of twelve. The dependency is not in dispute but the scoring axis is.

The American wall has no such condition. Protection, no performance test. A base that only ever sells to mandates and startups will be excellent inside the wall and helpless outside it. A wall without an export test breeds dependents. Korea’s bred Samsung.

Germany’s hidden champions never needed a wall at all. Their test was the world market from day one, and that is why they can pay $45 an hour.

The wall needs a door.

Builder Read

The opportunity is a filter, not a field. A gap is worth filling in America only when at least one is true:

• a mandate guarantees the demand (the 65% domestic content rule, the specialty metals clause);

• physics favors proximity (heavy, bulky, custom, fast-turn, what OSHCut‘s Caleb Chamberlain told me: “we beat them on speed, we tie them on heavy”);

• or labor is a small share of cost (a fab is mostly machines, which is why TSMC keeps doubling Arizona, now $265 billion committed).

Everything outside the filter stays imported, and should. Aim at the layer that passes the filter everywhere: the stuff that makes the stuff. The covered list is a parts catalog of things America now has to learn to build, with a date attached. What Survives to 2036 said pick narrow and go deep. This is the list to pick from.

I’m watching this

• Whether that actuator shop in Upstate New York ships in volume.

• And whether Chinese machine tools keep moving upmarket, because five domestic immersion scanners targeted for 2026 and 20 for 2027, against ASML’s roughly 130 a year, is a ratio today and a trajectory tomorrow.

The high-wage layer is reachable, but it is not reserved.

Const

No China this week. No robots. This week is about a question I’ve been circling for years, in other people’s companies and in my own: who owns your clock. Last week’s issue argued the empty trade is ten years. The fine print is that the ten years only work if they’re yours, and I’ve come to believe the way you fund a company decides exactly that.

Here's a number almost nobody believes. Venture capital funds about 4.4% of startup dollars. That's from the Kauffman Firm Survey of roughly 5,000 companies, and academic work (Da Rin, Hellmann & Puri) puts the share of firms that ever take VC below 1%. Even among the Inc. 5000, the fastest-growing companies in the country, a Kauffman survey found only 6.5% raised venture capital. The path every founder story glorifies is the rare one. The default American company is built the other way, and manufacturing is where the other way lives.

The Grind: Slow, Cheap, Painful, and Yours

The first way is the one I did. Start from zero, borrow as little as possible, win one customer at a time. In machining, the industry’s own data (NTMA, via Top Shops) says what that road looks like: the average shop under $10M revenue runs a 4.5% profit margin. Cross $10M and it becomes 8%, and the best cohort touches 15%. A $5M shop at 5% is worth maybe $750k after decades of work. Cross the valley to $10M at 8% and the multiple expands with you, call it $4.8M. Slow, cheap, painful, and at the end the thing is yours. Your clock, your customers, your name on the building.

I ran that company to profitability. And before it, two companies in the advertising space, in the very early days of the Internet, started from zero. Dry ramen in an office without a window. A complete nightmare of bookkeeping I had to run with no experience before I could even touch client work. The grind is real and so is the freedom. I think people who've never made payroll from cash flow underestimate how different a decision feels when nobody can call a board meeting about it.

The ceiling is real too. I wrote a whole issue about the $167k-per-employee trap, and the company that broke it, the sheet metal unicorn from last week ran exactly this sequence: eight years on cash flow, then one institutional round, roughly 10% sold at $1.01 billion, for capacity and not payroll. I won’t retell the story. The point for today is the order of operations.

And the pattern is much older than the current wave. Haas Automation, the company whose machines sit on half the shop floors in America, has been privately held since 1983. No rounds, no exit, one owner’s clock, and it grew into the largest machine tool builder in the country.

To me that’s the sequence that keeps the clock. Boring revenue first, capital second, and only for capacity.

You Can Buy Your Clock Back

The second way starts on someone else’s capital and steps off it. Software people write like they invented this move around 2017. Manufacturing did it decades earlier, at much higher stakes.

The most famous receipt in our industry: in 1981, thirteen Harley-Davidson executives, Willie G. Davidson among them, bought the company back from AMF for $80 million, almost all of it borrowed. The corporate parent had run the brand into the ground. The operators took the clock back, studied Honda’s Marysville plant, rebuilt manufacturing around just-in-time, and brought Harley to the public market in 1986 on their own terms.

Two years later, Jack Stack and 12 fellow managers bought a dying International Harvester remanufacturing plant in Springfield, Missouri, with 119 jobs on the line and almost nothing but debt. That became SRC Holdings and the whole open-book management movement. Buying your clock back is not a Silicon Valley invention. It’s an operator tradition.

The recent software cases just documented it better than anyone, with numbers, and they’re worth reading for the mechanics alone.

Buffer, 2018. Joel Gascoigne spent $3.3 million, about half the company’s cash, to buy out its main VC investors: 67.29% of the Series A shares, at an $80.8M valuation. The investors who sold made a 40.5% return over three and a half years. Roughly 10% a year, a savings-account outcome by venture math, and that modesty is exactly what made the deal possible. Gascoigne’s reason:

“It was clear that Buffer had become less of a fit for VC funding... I refused to compromise on the transparency and remote work aspects of our culture, so we started to explore slower growth goals.”

Wistia, 2017. The founders turned down an acquisition they describe as “a life changing amount of money” and instead took on $17.3M in debt to buy out their investors. Their logic is the cleanest statement of the treadmill problem I’ve found:

“Many people advised us to take on a new equity investor, but we knew with that strategy, we’d eventually be back in the same position: that investor would need a return one day too, just as our current investors did.”

Gumroad, 2017. The dark version. Sahil Lavingia’s own account: a $7M Series A from Kleiner Perkins, growth stalled, no Series B materialized, and he laid off 75% of the company to keep the product alive. Then came a transaction I still find hard to believe: Kleiner sold its stake back to Gumroad for $1, dropping the liquidation-preference overhang from $16.5M to $2.5M. Lavingia’s verdict on the whole structure:

“We were venture-funded, which was like playing a game of double-or-nothing. It’s euphoric when things are going your way, and suffocating when they’re not.”

The pattern holds all the way up. 37signals was profitable from day one, turned away nearly 30 VC firms (“all the VCs could offer us was cash and connections. We’re fine on both of those fronts”), and took exactly one check, a minority investment from Jeff Bezos personally, structured so none of it was needed to run the business.

I lived a version of this path with DigiFabster: raised VC early, then built toward profitability on revenue (and debt) when it became clear we couldn't raise anymore. It took almost six years to get to meaningful revenue, through COVID, inflation, and dried-up venture capital, all before manufacturing became sexy for investors. Almost 12 years of building. I kind of owned my destiny, but not entirely. Learned a ton.

The Treadmill, and What It Actually Returned

The third way is round after round, valuation as the scoreboard, exit as the plan. Manufacturing just ran a full experiment on this model, at scale, with the 2020-2021 vintage of digital manufacturing companies. Here’s what came back.

Fast Radius, the UPS-backed microfactory star, went public by SPAC in February 2022 at a valuation around $1.4B. Nine months later it filed Chapter 11. The assets sold to SyBridge for $15.9 million, less than ten cents on the dollar of what investors put in.

Shapeways, the original 3D-printing marketplace, SPAC’d in 2021 at roughly $600M. Chapter 7, July 2024.

Desktop Metal is the full arc in one company. SPAC at $2.5 billion in December 2020, market cap above $9 billion at the peak of the frenzy, acquired by Nano Dimension in April 2025 for $179.3 million, and 117 days later it filed Chapter 11 anyway. The binder-jetting IP and printer platforms, the crown jewels a $9B market once priced, liquidated for $7 million. Nine billion to seven million in four years.

Even the wins deserve a second look. Fictiv raised roughly $192M over a decade and sold to Japan’s MISUMI for $350M in 2025. A real exit, saved jobs, a strategic home. But an exit below 2x total capital raised flows through the liquidation-preference stack first, and what reaches common shares and founders after ten years is, by construction, a fraction of the headline. Xometry made it through the IPO window, and even there the founders kept 60% of the voting power through dual-class shares precisely because their economic slice was far smaller.

I want to be careful here. The graveyard doesn’t prove ambition is stupid. To me it proves something narrower: every 18-month round sells a slice of the company and resets the growth bar higher, and when the window shuts (mine says 2028-2030), companies built on that cadence have to shrink back to whatever their revenue per employee actually supports. Some can’t.

Some Businesses Can’t Be Bootstrapped, and It’s Better to Admit It

Now the honest half, because The Hopium doesn’t do strawmen. There are businesses where bootstrapping isn’t slow. It’s impossible.

Chris Power said it plainly on Core Memory this month:

“You need, I don’t know, over time, $30 billion, like maybe more than that.”

Elsewhere he’s framed the whole company as “$10B or zero.” That’s why he spends founder-hours on media and fundraising, and why I think the people mocking him for it miss the point: at that scale the capital requirement is the product requirement.

Kalanick’s Atoms just raised $1.7B before naming a single product, on the same logic. And the archetype for all of it, SpaceX, survived 2008 by a thread and needed two decades of billions before the Falcon 9 cadence looked inevitable.

Last year I wrote a post questioning exactly this approach. It went viral and collected close to a hundred opinions from manufacturers and investors, which tells you how real the risk is: nobody could agree whether this is the Tesla of machining or the next Hyperloop One. And still, I'm rooting for Chris. Honestly, I had my doubts then and I hold some now, even as a founder who's supposed to have the higher risk tolerance. But he's betting everything on his own conviction, and the industry needs people who make that bet. I respect that immensely.

The sharpest comment on that post came from Jason Premo, who runs a Swiss lathe shop and called it on the record: “prediction: flameout in 3-5 years.” Chris replied with his email address and an open door. I keep thinking about the fact that Premo’s window and mine, reached by completely different roads, land on the same years: 2028 to 2030. That’s when we find out.

I believe the test here is demand certainty, not ambition. If your customer is the Pentagon and your competition is onshore by default, the treadmill is the only road that gets there, and world-class fundraising genuinely belongs in the founder job description. If your customer is a purchasing manager who can switch to a cheaper quote on Tuesday, you are not Hadrian, and borrowing Hadrian’s playbook will bury you.

The Fourth Way: Buy What Somebody Already Built

There’s a path almost nobody glorifies, and the data on it is the most lopsided in this whole issue. McKinsey’s “great ownership transfer” report from February: by 2035, about six million American small and mid-size businesses face ownership transitions, including roughly 71,000 in manufacturing, with over a million viable sale candidates worth up to $5 trillion.

Now the number that stopped me. Under current patterns, 92% of exiting owners simply close the business. Only 5% manage a sale. Three decades of customers, process knowledge, trained people, and cash flow, defaulting to zero because nobody showed up to buy it.

That’s the fourth way: skip the twenty-year grind, buy one somebody else already finished, for a few times its cash flow, and bring the front-office playbook every issue of this newsletter describes. The crowd is chasing round three of a physical-AI startup. The empty room is a $2M-revenue shop with no succession plan, listed nowhere, about to close.

So…

For the wallet reading this.

Funding: underwrite the clock. Before the metrics, ask whether this company owns its time or rents it by the round, and whether its market actually requires the capital it’s raising.

Building: the sequence with receipts is the unicorn's. Boring revenue first, institutional money once, for capacity, on your terms. And if your wall genuinely needs billions, admit it on day one and treat fundraising as an engineering discipline, the way Power does.

Buying: 92% of retiring owners close instead of selling. I think that’s the single most inefficient market in American manufacturing, and the playbook to run what they built better is public.

One more thing about clocks, because I think it belongs here. This week I read a post celebrating billionaires who still work ten-hour days in retirement, offered as proof that people are meant to work forever. A commenter named Ken pushed back with a line from his mother:

“You are going to spend the rest of your life inside your own head, so make sure you decorate it with beautiful things.”

I’m on Ken’s side. Work gives purpose, and I love this industry, but it shouldn’t be the sole purpose that eats every hour you have. And I believe the funding paths decide this too.

The bootstrapper can take a Tuesday for his kid’s game and answer to nobody. The founder on the treadmill checks with the board. Small difference in a week, enormous over ten years. So when I say the ten years should be yours, I mean all of it: the company, the hours, and the choice of what fills them.

I’m Watching This

1. Hadrian’s promised 2-million-square-foot factory, “the biggest buildout outside Elon’s Gigafactory” per Chris Power, and whether any audited number ever appears behind his “billion plus in revenue next year.” The $1.6B January round was led by T. Rowe Price, which is pre-IPO money. The treadmill’s biggest live bet in our industry, running in exactly the window my 2036 issue says decides everything.

2. The CMMC Reform Task Force. DoD paused third-party certification in July and promised findings within 60 days of its RFI. Thousands of small shops spent six figures getting certified because the government asked them to. If reform kills the third-party requirement, that compliance moat evaporates and the committed get punished for committing. Watching what survives.

3. Atoms' $1.7B war chest meeting its first real product. Kalanick merged everything he owns into one company and named mining, construction, and food production as targets, all sectors with no federal anchor. The biggest live test of whether a bits founder can buy through the atoms wall.

Last week the empty trade was ten years. The fine print: the ten years only compound if the clock is yours.

Const

Last week’s 2036 issue argued the wage never cleared the market. This week my good friend Mike Payne named the metric that proves it, a unicorn showed the fix, and a 20-year robotics retrospective supplied the clock. I’ve come to believe this one number decides who survives the window, so I’m giving it the whole issue: what a US factory has to look like to still be here after the money leaves.

The Number That Matters More Than Headcount

Mike runs Hill Manufacturing, hosts Buy the Numbers (a podcast about the actual financials of running a shop), and buys machine shops the way I collect signals. After the 2036 issue he wrote something I’ve been chewing on all week:

“We keep telling ourselves it’s a labor shortage. Kids don’t want to work. Nobody respects the trades. Honestly, it’s a lazy excuse... Here’s the number that matters more than headcount: revenue per employee.”

And the test that goes with it:

“Can you pay a welder six figures and still make money on the job. If the answer’s no, you don’t have a hiring problem. You have a pricing problem.”

Run the math and his point stops being an opinion. A $100k welder costs roughly $130k fully loaded (payroll tax, insurance, benefits; estimate, not gospel). Direct labor in machining and fab typically runs 25 to 35% of revenue. So paying six figures comfortably requires somewhere around $400,000 to $520,000 of revenue per employee.

Now the industry’s actual number. Mike, who buys shops and looks at their books for sport, described the average American precision machine shop on his own show:

Roughly 18,000 to 19,000* job shops in the country, and the average one runs “about 10 machines, 12 employees, 2 million in sales.”

* Personally, I’m a bit more skeptical about the actual number of job shops in the US. My best guess is it’s closer to 10,000.

Twelve people, $2 million. That’s $167,000 of revenue per employee.

There’s the whole labor crisis in two numbers. The six-figure welder needs a $450k-per-person business. The average shop is a $167k-per-person business. The wage can’t clear because the revenue per person can’t carry it, and no job fair, no trade-school campaign, no culture war about kids these days changes that equation by a dollar. To me, the house cleaner from the 2036 comments (earning $20 to 30 an hour, offered $13 to 17 in every trade she priced) read manufacturing’s P&L more honestly than most of its owners do.

And Mike’s diagnosis of WHY the average shop is stuck at $167k:

“The owner has become the bottleneck... they do the sales, they do the quoting, they’re scheduling the floor, making sure invoices went out.”

I spent 12 years watching exactly this person: the owner quoting at midnight after running the floor all day. I built a whole company for them. The shop floor was never the problem.

The Unicorn Ran Exactly This Play

Which brings us to SendCutSend (again, I know). The raise itself is a month old (mid-May: $110 million at a $1.01 billion valuation from Sequoia, Paradigm, and Stripe’s Collison brothers), but The Fabricator’s deep dive published this week finally put real numbers under it, and to me the numbers are the story. A sheet metal shop from Reno, valued like a software company, because in the way that matters it is one.

The numbers underneath: revenue “well over $150 million,” about 450 employees across four states. That’s $333,000 per employee and climbing, double the industry average, on the road to the six-figure-wage threshold. And here’s the detail I think many will miss while staring at the valuation: SendCutSend got there with a largely manual floor. Manual denesting. Manual bending. Operators with suction lifts. No lights-out automation anywhere.

What they automated instead was the front office. Upload a file, get an instant price, click accept, and the order nests itself and queues to a laser in minutes. One order was nested and on the laser minutes after the customer clicked accept, and boxed 27 minutes later. Their lasers run at 88% beam-on time, a number most shops don’t hit in their dreams, not because the machines are magic but because no order ever waits on a human quote, a human nest, or a human schedule. The late industry consultant Dick Kallage said it decades ago:

“You’re not in the metal fabrication business. You’re in the information processing and changeover business.”

SendCutSend is what a shop looks like when someone finally believes him.

The labor model completes the picture. They hire from “the donut shop, Dollar Tree, and Walmart” and train in-house, from day one, by design. High revenue per employee doesn’t require finding skilled workers who don’t exist. It requires a system productive enough to justify manufacturing the skill yourself.

Founder Jim Belosic’s own account of the raise is the part I reread more than once. He funded the company on cash flow for eight years, stayed profitable, and then:

“We held out until the ‘reindustrialize’ vibe shifted and we found partners who actually understand that hardware is hard.”

An operator built boring revenue first, watched the capital window open, and priced it on his terms: sold 10%, one raise,

“I don't want to be on a treadmill, out there raising funds every 18 months.”

That’s the sequence. Revenue per employee first, venture money second, and only for capacity, not payroll.

Even his competitor endorses the logic. Caleb Chamberlain, who runs OSH Cut and has every incentive to dunk, wrote instead that the valuation reflects a real bet: web-first players still account for less than a thousandth of US manufacturing output, McKinsey counts 71,000 manufacturing owners exiting by 2035 with over a quarter unsure of succession, and somebody’s systems will absorb that demand. Mike Payne sees the same wave from the buyer’s chair:

“I show up, talk to some shop owner and they haven’t prepped, they haven’t done anything. They’re just ready to be done today.”

Seventy-one thousand $167k-per-person businesses are heading for the exit. And the question is what replaces them.

The Ten-Year Clock

Here’s where the impatient money should pay attention. Max Lobovsky, who co-founded Formlabs, posted a 20-year robotics retrospective this week that every reindustrialization deck should be forced to include. Collaborative robot arms: 40 years in, roughly 8% compound annual growth. Autonomous vehicles: the single biggest success of the cohort, and it took the full 20 years and around $100 billion to reach limited commercialization. Humanoids: two decades after Robonaut, still mostly “incredible demos” and entertainment. His conclusion:

“Impressive demos usually precede meaningful commercialization by a decade or more. If you want to build something that scales in the next decade, look for an idea that was demonstrated 10+ years ago.”

Apply that filter to this issue and it clicks into place, because web-first manufacturing is even older than most people remember. Quickparts was selling instant quotes for custom parts back in 1999 (3D Systems bought it in 2011), and Protolabs was quoting molded parts with software in the early 2000s when it was still called Protomold. 247TailorSteel put the full model on a European shop floor in 2007 (today: 70,000 parts a day).

SendCutSend and OSH Cut proved the US sheet metal version in 2018. So the unicorn is a quarter-century-old idea finishing its apprenticeship, which by Lobovsky’s rule makes it exactly the thing that scales this decade.

I lived a piece of this timeline myself: I spent a decade pitching instant quoting to shop owners and hearing “our parts are too complex for that.” The tech was never early. The industry was late. The robots that will eventually automate these floors run on the same clock, which is why the smart operators automated the inbox first and are only now, with the capital and the order book, buying the robotic brakes.

Now… the uncomfortable half of the lesson, because the same clock convicts the opposite strategy. Venture money arriving this fast into hardware tempts founders into the oldest trap in the window: hire ahead of revenue. Headcount is the most photographable form of traction and the most expensive form of payroll, and when the window shuts (my 2036 issue put that around 2028-2030), a company built to 200 people on raised money has to lay its way back to whatever its revenue per employee actually supports. In manufacturing, where trust is the asset and communities are small, that’s often fatal. The companies that survive the window are the ones that treat it the way Belosic did: a chance to buy machines and capacity, never a chance to buy headcount that revenue hasn’t earned yet.

And the clock explains one more thing I keep seeing. A thread going around this week on X described Gen Z entrepreneurship as "running a portfolio of asymmetric coups": an AI wrapper one month, a TikTok Shop product the next, "extract as much as possible while the window is open, then abandon it as soon as the trade stops working."His best line:

"The old founder wanted to build one company that would outlive him. The new one wants enough wins that he never has to depend on one company at all."

I won't pretend that's stupid. It's rational, and that's the point. Run it against Lobovsky's clock and manufacturing becomes the one arena the day traders will always self-select out of, because nothing here pays inside a rotation cycle. Commitment looking irrational everywhere else is exactly what makes it the edge here. The crowded trade is optionality. The empty trade is ten years.

Zoom all the way out and this is also the only honest answer to the China question.

The US runs roughly $2.9 trillion of manufacturing value added on about 12.9 million workers, call it $220,000-plus per worker. China runs $4.9 trillion on a manufacturing workforce so large it’s genuinely hard to count, putting its number somewhere around $40,000 to $50,000 per worker. America is already four to five times more productive per person, and it cannot win, will never win, a headcount contest against a country with four times its industrial workforce and five-year plans against our four-year attention spans.

I’ve walked Shenzhen’s shop floors (30 to 90 person operations, three shifts, running 24/7), and I can tell you there is no American Shenzhen coming, not overnight and not this decade. The one that eventually emerges won’t look like Shenzhen at all. It will look like SendCutSend’s Reno: smaller headcount, higher automation, $400k-plus per person, wages that finally clear. Fewer people, much more per person. I believe that’s the only version of reindustrialization that can actually work, and I’d rather build for it than mourn the 1979 photo.

So…

For the wallet reading this. Funding: underwrite revenue-per-employee trajectory, not headcount growth; in this sector, hiring ahead of revenue is a liability wearing a traction costume.

Building: track revenue per employee monthly and margin per FTE if your mix is wide (Mike’s refinement); price so a six-figure wage works or fire the customer that forbids it; automate the inbox before the floor, because the unicorn’s first $150 million came from a manual floor and an automated front office.

Buying: 71,000 owners are exiting $167k-per-person businesses that trade at 2 to 4x EBITDA, and the playbook to double their RPE is now public. That’s the least glamorous, highest-certainty trade of the entire buildout.

I’m Watching This

1. Musk quietly bought APR Energy: ~$1B, revealed only by an SEC filing, for 1+ gigawatt of truck-mounted turbines that skip the five-year grid queue. Every AI lab can buy the same chips; one now owns power that ships by truck. Watching for the first follower, because one buyer is an anecdote and two is the co-location thesis going mainstream.

2. OSH Cut is opening a second factory near Cincinnati. The web-first expansion continuing under the unicorn headline, without the venture money.

3. Xometry Q1 revenue up 36% year over year. The marketplace end of the same model, growing at public-market scale.

The buildout doesn’t need more people. It needs more per person.

Const

This issue is different from the usual. No single news story. I took the signals I’ve tracked all year, checked the numbers, and ran the tape forward to 2036. Rules of the experiment: no world war, no collapse, no miracle technology that doesn’t already exist in a factory somewhere. Just today’s slopes, extended. Where I’m guessing, I say so.

One more input this week: the roughly 200 comments you left on the jobs issue. A few of you wrote the middle of this piece for me. Keep doing that, please.

Ten-Year Money, Four-Year Politics, Five-Year Plans

Start with the least romantic chart in the buildout. US manufacturing construction spending peaked in 2024 at $235.6 billion annualized and has been sliding since: roughly $206B in November 2025, $190B by March 2026. Down about 8% in four months. Not a crash, but a slow leak. The concrete-pouring phase of reindustrialization peaked two years ago.

Now the other chart. The hyperscalers plan roughly $700 billion of AI capex in 2026 alone: one year of datacenter spending, about 3.5x the entire construction run-rate of American manufacturing. The country is running its largest infrastructure buildout since the interstate highways, and it’s not for factories. Mills and smelters now bid for electrons against trillion-dollar balance sheets, and on a legacy power contract the mill loses that auction every time.

My good friend Zane Hengsperger wrote a factory-abundance manifesto last week (read it, he’s a builder, and his infrastructure numbers check out against USGS to the decimal) and buried his best idea in the middle: energy-intensive industry has to stop being a grid customer and become a grid asset. Generation on-site, load shed back when datacenters spike, sites co-located around shared interconnection because the queue for a new hookup is measured in years. Hold that idea. It comes back at the end.

Then the third clock: venture. Defense tech hit an all-time funding record this year, $14.6B into autonomous-weapons startups in the first half of 2026, and Fortune is already asking whether it’s a bubble. In the sense that matters, it is: a venture fund runs on a ten-year clock, an administration on a four-year one, and the money flooding hardware needs exits by the early 2030s.

Will Alverson, another builder I trust, put the real game plainly: a big part of reindustrialization is simply outlasting China’s ability to book consistent losses. He’s right, and that’s the tension. Nobody knows how long outlasting takes, but every precedent of winning through booked losses (DRAM in the 80s, solar in the 2010s) took the winner a decade-plus, so call ten years the floor with no known ceiling. We fund that with 10-year money and govern it in 4-year increments, against a country that publishes its industrial priorities in five-year plans and means them.

So my first 2036 call, the one I hold with most confidence: the tourist capital leaves around 2028-2030. Not because reindustrialization failed, but because it stopped photographing well. VC into manufacturing reverts to its historical trickle and the word “reindustrialization” retires the way “web3” did. What survives is three kinds of company:

1. Whoever has a government check stapled to the balance sheet (the Tier 2 pattern from my last issue, where every serious rare-earth entrant carries a federal commitment)

2. Whoever reached boring revenue before the window shut (parts, MRO, aftermarket, contract assembly)

3. And whoever supplies the one hardware buildout that keeps its own money printer: robots.

And the “a big war will fix it” theory: I don’t buy it, especially as a plan. The next decade looks like this decade, regional conflicts where big powers test hardware through partners. Ukraine is on track to field 3 million FPV drones in 2025, and last week they confirmed testing underway on a US-Ukraine drone production venture valued at $35 to 50 billion, up to 200 Ukrainian companies, plus a track licensing Ukraine to build Patriot interceptors at home. Look at the direction of flow: America buys battle-proofed drone designs from a partner at war, the partner gets a license to make American air defense domestically. Knowledge migrates to wherever production actually happens. The world stayed connected, logistics still work, and any 2036 scenario built on wartime autarky is fan fiction.

The Vault Flipped

Second thread, and to me the deepest one. The frame everyone over 40 still carries (“China steals IP”) is dead. It died quietly in the last few years, and this week produced the cleanest evidence yet of what replaced it.

China now defends process IP the way the US defended it in the 1980s. CATL licenses its LFP battery chemistry to Ford’s Kentucky plant, reportedly at around 10% of production value, with zero equity and zero board seats. Ford builds the plant, employs the workers, takes the capex risk, and pays rent on the knowledge.

And Kentucky is one outpost of a bigger play. Blocked or tariffed out of rich markets, Chinese automakers build inside them instead, country by country: factories in Hungary, Indonesia, each one turning a host government into a stakeholder in a Chinese company's success. Analysts call it industrial diplomacy. By one estimate, Chinese EV and battery investment abroad now outpaces US firms 4 to 6x.

This month Reuters reported Beijing is drafting tiered export controls on its own frontier AI models. China is building export controls for intelligence, the same tool Washington built for chips. The thief became the vault.

Meanwhile the US version of IP protection played out in a courtroom filing on July 10. Apple sued OpenAI, alleging a departing engineer walked out with a 1,000-plus page compilation of main-logic-board manufacturing and testing documentation for Jony Ive’s device program. The texts in the filing are almost too on-the-nose (”LOL, I found out I can access the network storage, so funny”). Strip the drama and look at what was worth stealing: not model weights, not chip designs. Manufacturing and testing documentation. The boring, decades-accumulated knowledge of how you build reliable hardware at scale. That’s the asset now. Both superpowers told you so in the same week, one with export controls, one with a lawsuit. And containment leaks the other way too: the same week, OpenAI and Google were reported selling AI services to Singapore subsidiaries of Alibaba, Baidu and Tencent, all three on the Pentagon's 1260H list, legally, via the shell - American corporations already run cheap Chinese open-weight models for the same reason American consumers buy Chinese hardware on Amazon: almost as good, never more expensive.

Watch the space race through this lens. On July 10, CALT caught a Long March-10B booster with a net on a sea platform. Not landing legs. A net and four hooks, a genuinely different recovery architecture. That’s what isolation does: China stopped copying and started engineering around, and the result is proprietary process IP that no export control can claw back. The gap that remains is cadence, not capability. SpaceX reflies boosters in under 30 days; CASC hopes to refly this stage before the year ends; China ran about 70 orbital launches in 2025 against SpaceX’s 130-plus. By 2036 I expect two fully separate space-industrial stacks, no shared supply chain anywhere in either one, with the US ahead on reuse cadence and China ahead on the thing underneath: the manufacturing base feeding the pads. Not Soviets-vs-US, where one side’s economy was hollow. Both economies are real this time. That’s what makes it a new kind of race.

So the 2036 picture on IP, my second call, medium confidence: theft stops being the story and licensing becomes it, with the rent mostly running east to west. More US plants operating on licensed Chinese process knowledge, the Ford Kentucky structure repeated across batteries, magnets, materials processing. This is not a doom call. Licensing in was exactly how Japan, Korea and China themselves climbed. The doom version happens only if the US pays the rent and never develops operating knowledge of its own. Which brings us to who actually develops process knowledge: people who run production, not people who fund it.

Nobody Wants The Jobs, Still

Third thread, the human one, and the place where I want to kill the romance completely.

The culture argument says America’s bottleneck is status: make the machinist the main character again and the labor pipeline fixes itself. The problem with this theory is China. Young Chinese don’t want the CNC job either. China’s flexible-employment bucket (gig work, informal work) is expected to hit 320 million people this year, somewhere around 40 to 44% of the workforce depending on whose denominator you trust, and Xi is personally telling kids to study manufacturing because the factories can’t hire. The country that supposedly out-cultured us on manufacturing is watching its own kids choose food delivery over the shop floor, because the shop floor doesn’t pay a city rent. This is not an American culture problem. It’s a rich-country pay problem, and China got rich.

Then my own comment section did the reporting for me. The jobs issue pulled about 200+ replies, and the most useful ones didn’t come from operators or investors. They came from the people the buildout says it can’t find.

Cabri Chamberlin, a college graduate cleaning houses for a living, wrote the whole labor market in two sentences:

“There is no shortage of workers willing to do these jobs. There’s a shortage of pay and dignity in those fields.”

She priced the pivot into a trade the way a CFO prices capex, because for her it is capex: a phlebotomy certification would cost thousands of dollars and six months of her life, for starting pay of $13 to 17 an hour in her area. Cleaning houses pays her $20 to 30. Every trade she looked into ran the same math. Her verdict:

“Until the pay and conditions of the skilled trades rises above glorified indentured servitude, the rational choice is to piece together gig work and ‘unskilled’ part time jobs.”

Call that laziness only if you can't subtract.

Another reader, Gavin, pointed me at the chart I should have led with: male labor force participation (FRED LNS11300001), down from about 86% in 1950 to roughly 68% today. One unbroken 75-year slope, tens of millions of men neither working nor looking. His line:

“America has plenty of workers.”

It does. What it doesn’t have is a wage that clears the market, and a third reader (Citizen Deux) named the clearing price: half a million skilled workers needed in electrical, welding and pipe fitting, and

“Opening salaries are going to have to be north of 100k.”

Gavin also named the game theory that keeps both wages and training stuck: spend $10,000 training welders and your competitor spends $6,000 poaching them. You go bankrupt, and then his pipeline dies too. Nobody trains because everybody poaches.

Ed Tate on X pushed back on that trap with the old fix, and it’s worth hearing: training expenses tied to continued employment. His first employer ran scholarships where you worked day-for-day for the time they covered, repayment owed if you left early.

“Employees leaving prematurely was never an issue.”

Unions ran the same play for apprentices. The mechanism is real and it worked for decades. But notice what it requires: an employer confident enough in its five-year order book to front the money, and pay good enough that signing the bond feels like a deal instead of a debt trap. Get the pay wrong and a training clawback is just Cabri’s “indentured servitude” line with a signature on it. Get the order book wrong and no CFO approves the program in the first place. The contract is the easy part, but the certainty is the hard part.

And Tom Bomer, retired in 2013 after 46 working years (33 of them building flight simulators), closed the loop from the far end of the age curve:

“There must be training programs that are well run, and that pay, from the beginning, a wage that makes people understand that their skills will be rewarded for years to come.”

Rewarded for years to come. That’s demand certainty, seen from the worker’s side of the table.

The range was wide, to be fair. A retired shop owner told the young to do the work and quit the bitching. A reader in Mexico said his country’s factories have no hiring problem at all, which is our rich-country thesis reporting in from the other side of the curve. Australians described the identical disease. Every side of the culture argument showed up, and not one person argued the pay was fine. When 200 people disagree about everything except one thing, that one thing is the story.

Which brings me to Willow Run, because everyone in this movement eventually reaches for it and almost everyone grabs the wrong end. Watch the reconstruction (thanks to Paul van Metre for posting this) of how Ford built a B-24 every 63 minutes by 1944, 428 bombers in the peak month, 3.5 million square feet, 42,000 workers.

The memorable part is the line. The instructive part is everything around it. Demand certainty built that factory, not culture: Roosevelt asked for 50,000 aircraft, the government bought everything the line could produce, and Washington built a freeway (today’s I-94) to move the workers. An order book that certain is also the only known solvent for Gavin’s training trap, and Ed Tate’s contracts scale under exactly those conditions: Ford could train tens of thousands from scratch because no competitor could poach against demand the government had already bought. A third of those workers were women earning the same 85 cents an hour as the men, and they came because the pay was real, not because riveting got glamorous. Status followed the paycheck.

It took two years of public failure before the line worked (37 bombers in the first month, the whole state joking “Will It Run”), it required converting the largest industrial machine on earth (Detroit), and it shut down ten months after peak, the day the customer stopped buying. The greatest factory in American history was a demand artifact. The US of 2026 has no Detroit to convert and no FDR-scale customer except one, the Pentagon, which is why every durable company in this movement already has a federal anchor. And that’s not a scandal. That’s the playbook working as designed, just smaller.

So what does the floor actually look like in 2036? Fewer humans, paid properly, supervising fleets. Tesla is converting capacity toward Optimus with a reported 2026 target of 50,000-100,000 units against roughly 10,000 unique parts each, and around 70% of that supply chain is Chinese today: the reported $685M actuator order went to Sanhua, enough for about 180,000 robots, while US production plans fuel the rumor mill. Unitree already ships humanoids profitably at $4,290. I stand by my tier call from last issue: the US loses the cheap-volume robot race and shouldn’t enter it. But every robot deployed on a US floor, whoever built it, changes the labor math the same way: the job stops being “operate the machine” and becomes “own the cell.”

My third 2036 call, medium-high confidence: US manufacturing employment stays roughly flat even as output rises, but the median floor wage rises meaningfully, because the person left standing runs six machines and a robot fleet, and there’s no version of that job that pays badly. That’s the only honest path to Citizen Deux’s north-of-100k number: not status campaigns, revenue per worker high enough to pay it. Cabri doesn’t become a welder at $17 an hour. Either the cell pays her six figures to run it, or she keeps cleaning houses and she’s right to.

And the software those floors run? Jack Bookey (Nox Metals) posted the most useful number I’ve seen on it: assume 80% fidelity per human handoff and only 26% of a signal survives the round trip from company brain to factory floor and back. Nox’s answer is one schema spanning quote, ERP and edge stations on the floor. Software shaped exactly to their plant, in other words. And Caleb (OSH Cut) reminded me publicly last week that even “solved” parts aren’t solved: the on-demand space covers the easy stuff, the breadth of real capability barely at all.

Put those together and here’s my honest 2036 software read. The system of record doesn’t go away; every plant still needs one, and consolidation means fewer, deeper vendors. What changes is where the value sits. The fight moves up a layer, to the intelligence that watches the floor and closes Bookey’s fidelity gap, and that layer gets built two ways: in-house by the operator class that can staff it (Nox, OSH Cut, SCS, and many others now), and as products for the quarter-million US manufacturers that never will.

The ERPs that open themselves up to that layer become more valuable, not less. The ones that bolt a chatbot onto 2015 architecture and call it intelligence get found out. Either way, the largest software prize in American manufacturing is still unclaimed as I write this.

The 2036 Picture

Pulling the threads into one sketch. Speculation, held loosely, ranked by confidence.

High. The steel gap doesn’t close (79.5 Mt vs China’s 1,005 Mt is not a ten-year fix) and chasing the aggregate isn’t the point; new mills and smelters rise only where the energy math closes, on-site generation, flexible load, datacenter co-location. Zane’s grid-asset idea stops being clever and becomes the default site plan. And training-with-strings returns: as wages climb toward the clearing price, Ed Tate’s scholarship-and-clawback contracts become standard hiring architecture in the trades.

Medium. The US keeps two or three genuine process-IP strongholds (reusable launch, frontier fabs, possibly biomanufacturing) and defends them with China’s own export-control playbook. Two fully separate space stacks, cadence advantage US, industrial base advantage China. Humanoids (or some sort of hybrid) everywhere on American floors with a BOM still 40-70% Chinese, unless robot builders spend the decade qualifying domestic suppliers: at 100k units and a $20k BOM, the non-China share is ~$600M a year of work looking for shops. The biggest open bet in the picture.

Low, the wildcard. Demographics. China’s curve is steeper: population shrinking three years running, fertility near 1.0, the 320-million gig number hiding how thin the formal pipeline is. America historically replenishes its workforce through immigration; China doesn’t. Whether the US keeps that advantage is a policy fight I won’t touch here. I don’t know how this resolves. Neither does anyone selling you certainty.

The shape holds either way: 2036 US manufacturing is smaller in headcount, higher in wage, radically more automated, clustered in defense- and energy-anchored islands of density, running software its owners control, paying licensing rent on some process knowledge and collecting it on the rest.

Not the 1944 photo. Not the 1970s employment chart. Smaller than the speeches promise, better paid than either.

So…

The window is the strategy.

Funding: assume exits into a post-tourist market and underwrite to revenue, not narrative; a federal anchor is a feature, not a stigma.

Building: boring revenue before 2029 or get that anchor, and treat the robot BOM as the largest supplier-qualification opening since automotive (actuators, gears, harmonic reducers, test and pack, currently shipping from a country your customer is being told to diversify away from).

Supplying: the co-location wave is where new-site engineering money goes next.

Hiring: forget the millions who aren't coming. Design cells for the six people who are, pay them like they run the place (they will), and put an Ed Tate contract behind the training.

I’m Watching This

1. Does the Long March-10B booster actually refly before the end of 2026? Cadence, not capability, is the tell on how far behind China’s reuse really is.

2. Do China’s frontier-model export controls get enacted, and with what scope? Drafting is a signal. Enactment makes AI decoupling official in both directions.

3. Does manufacturing construction find a floor above ~$150B annualized? That’s the line between “the buildout matured” and “the buildout was a sugar high.”

The buildout is real. The window is short. Those are not the same sentence.

Const

The $1.77 trillion in announced US manufacturing commitments since 2025 is real. The problem is where it's going. Last few weeks I dug into which products carry a US cap table and a Chinese BOM, who gets funded, who staffs the factories.

So this week I tried to map the actual picture: what’s worth building here and what isn’t, what pulls capital and what doesn’t, and where the real gaps are. This is the strategy layer, chapter four of four.

The Filter

I believe two questions do most of the work here. One, if China cut this supply tomorrow, how bad is it, across how many sectors, how fast. Two, can the US actually compete, given our power costs, our labor, our patience with capital. Put those on a grid and you get four tiers. Most of the money is piling into the tier that already has money.

Tier 1: Build Here. Nobody Is.

High pain if it’s cut, and the US can genuinely compete. The gap is open because capital showed up wrong, or didn’t show up at all.

The clearest one I’ve found is on-demand FATP. If you don’t live in hardware: FATP is final assembly, test and pack. The last mile where someone takes your pile of parts, screws them together, checks it works, and boxes it. What Foxconn does for Apple. SendCutSend and Xometry already made parts a solved problem (ten ways to get a bracket cut by Tuesday). But assembly… nobody solved.

Jabil runs 50,000-unit programs, not 500-unit runs with the boards, the mechanical, the test and the pack in six weeks at a price a seed-stage startup can pay. So every hardware company still flies to Shenzhen by unit 200. em m0houris, a hardware operator, put it plainly on X:

“we have a glut of part production, but the quiet part no one is saying out loud is that someone needs to put it together. this is the hard part.”

I went looking for the SendCutSend of assembly and found no pure play, only companies circling it. Bright Machines raised $126M to automate assembly cells. Formic and Tutor Intelligence rent robot labor by the hour. MakeMatter is building autonomous contract manufacturing for robotics, and RCM does domestic PCB assembly at small scale. But a service where you ship 400 units of design and get a finished, tested product back? Didn’t find it. If you know one, reply in comments, I’ll feature it. Real ask, not rhetorical.

The one that actually gets me is nitrile gloves. The US imports 99% of them, and the raw rubber (NBR) traces back to China almost no matter which country dips the final glove. Scott Maier built a fully domestic plant in Wytheville, Virginia, got it 80 to 85% complete, and the Development Finance Corporation, the exact tool built for this, issued zero loans under its PPE mandate. The equipment is being sold off this year. His own words:

“USG does not see the need to support a domestic nitrile glove industry”

A $0.03 part (!) that stops chip fabs, pharma, food and defense at the same moment if it dries up, and we’re selling the plant for scrap.

Tier 2: Build Here, But Only With Government

High pain if it’s cut, but the unit economics never close on their own. My read: the market is never funding this alone, and you can see it in who shows up. The government, every time.

Rare earths are the cleanest example. China refines about 90% of the world’s supply. MP Materials runs the only US mine and processing site at Mountain Pass, and look at how it scaled: the DoD became its largest shareholder, $400M in preferred stock, plus a $110/kg price floor. That’s not your Series B-C. That’s a government guaranteeing the price so the plant survives. Everyone else in the lane looks the same. Phoenix Tailings, a VC-backed refiner, has $116.6M in venture money plus a Pentagon commitment up to $500M. ReElement, an Indiana refiner pulling rare earths out of scrap magnets with chromatography, carries an $80M Pentagon loan (under review, they have doubts it can scale). Energy Fuels, an old uranium miner, just landed a $725M government loan and is spending it on a century-old German magnet maker. Every serious entrant has a government check stapled to it. That’s the tier telling you what it needs.

Gallium and germanium are worse. China makes about 98% of refined supply, restricted exports in July 2023, banned them to the US, then temporarily suspended the ban through November 2026. US primary gallium production is basically zero, and I found no funded US company fixing it at scale. Another open gap, Tier-2 flavor. It needs a strategic reserve, not a pitch deck.

Same story with aluminum: 23 US smelters in the mid-90s, four today, because primary aluminum needs enormous, cheap, uninterrupted power and nobody builds a smelter without a decade of subsidized electricity behind it.

What I actually want is a directory for this tier. Not another list of VC startups building the shiny stuff, but the raw dependencies nobody here will touch: the materials, the precursors, the processing that independence actually runs on, sorted by category and by how badly it hurts if it’s cut. Most of it looks unrealistic or just too hard, which is exactly why it sits open.

Some of it is starting. Musk put over $1B into a lithium refinery outside Corpus Christi that began producing in 2025, the first large-scale battery-grade lithium refinery in the US, pointed straight at the China processing dependency. That’s one plant. We need a hundred, and a map of where they go.

Tier 3: Leave It To China

Low pain, or China’s lead is just permanent. Money spent trying to win it back is money wasted. Here are two examples.

Consumer electronics glass. Lens Technology, run by Zhou Qunfei (mentioned in one of my previous Hopium’s), makes the cover glass on basically every iPhone, and now display glass for 30 car brands including Tesla. She started grinding watch glass on an assembly line at 16 and spent 20 years building that base. Corning still makes the raw glass here, but the fabrication and finishing at that scale and margin is China’s, and no US plant is winning it back. That’s a leave-it, not a fight.

Humanoids, I want to be careful with, because it’s not as simple as “China won”. The US has the deepest bench at the frontier. Figure is deploying at BMW, Agility’s Digit is doing real, paid warehouse work, and Apptronik, Optimus, 1X and Boston Dynamics are all serious on capability and embodied AI. That’s the American strength: new capability, not cheap volume. But the numbers are brutal. Unitree shipped 5,500 humanoids last year, profitable, cheapest unit $4,290. Agility’s Digit runs about $250k. Tesla is targeting $20 to 30k and hasn’t shipped one commercially.

So my read: short term, racing China on cheap volume is a losing bet, leave it. Long term, if Tesla converts Fremont to a million units a year or Figure’s line hits target, the gap narrows and it changes. So build humanoids, just don’t try to win them on the price of the unit. Win on the capability and the software, and let China make the cheap hardware.

Tier 4: The Complicated Middle

Real security value and a weak commercial case. Hadrian ($1.6B valuation, Founders Fund and Lux) is building CNC capacity for defense, the right bet, though parts were never the only bottleneck. Its CEO Chris Power named the real wall:

“there aren’t enough skilled tradespeople to even be hired. We need millions and millions and millions”

Chips are funded right, TSMC putting $165B into Arizona with $6.6B of CHIPS money behind it. But… the question that is rarely asked: who assembles the finished systems at volume? FATP again, one tier up.

Where The Money Actually Goes, And Who’s Quietly Doing The Work

Look at Relay’s map of the buildout: 525 companies, 134 in space, 111 in defense, 96 in manufacturing. That’s where venture clusters, the part of the stack that photographs well and fits a 10x story.

The part people underrate is who’s grinding on the unglamorous gaps with no venture money at all. Russell Winter keeps a manufacturing roll call on X, a couple hundred small US shops describing themselves in five words: cast iron foundries, PCB assembly, spiral bevel gears, Swiss lathe work, domestic contract assembly. Caleb Chamberlain (OSHCut) is in it. These are the people actually playing the long game, and almost none are fundable by a fund.

There’s also a second wave forming right under the Tier-1 names: shops setting up specifically to supply them. Gears and harmonic reducers for the robot builders (a part the US mostly imports from Japan and China today). Re-engineered OEM replacement parts on faster lead times. The small, boring subsystems every Figure or Anduril will need second-sourced at home. That is exactly how China’s supplier density got built, one specialized shop at a time. Watching that layer form is more interesting to me than the next $300M defense round. Gavin (@67Designs) keeps saying it:

“We are doing this to ourselves. Stop blaming China. Start playing the long game.”

I think he’s right. The grid above just tells you which parts are worth playing.

So…

If you build one thing off this issue, build on-demand FATP. Most undercapitalized white space in US hardware, and whoever builds the Jabil-for-400-units owns the chokepoint every hardware startup hits on its first real run.

Rare earth processing is real too, but it’s an infrastructure play, not a startup: get the offtake or the price floor signed before you break ground, the way MP, Phoenix Tailings and Energy Fuels did.

And if you’re an operator in that roll call, or setting up to supply the Tier-1 primes, the gap map above is your addressable market.

Everything in the “leave it to China” box is a trap that burns capital to prove a point.

I’m Watching This

• Defense manufacturing as the Pentagon’s own bottleneck. Allison Giddens, a manufacturing exec, summed up the irony on LinkedIn: parts sit ready for eight weeks while the government’s paperwork holds up the shipment. The constraint Washington blames on industry is partly Washington.

• The $73B missile budget hitting a propellant wall. The whole missile-defense buildup runs through two solid rocket motor makers. The 2027 goal is roughly 5,000 interceptors a year, projected deliveries sit above 2,100. Emergency cash doesn’t fix it, because you cannot surge artisanal propellant.

• China’s AI playbook, said out loud. The argument going around: give away world-class models for free, export cheap inference on cheap electricity, make the world depend on your models, chips and standards. Profit optional, power isn’t. If that’s the game, “but they’re losing money on it” is not the rebuttal people think it is.

Till the next week…

Const

I’m Const. 12 years in US manufacturing. Ran a machine shop, built DigiFabster (quick quoting for job shops), worked with hundreds of manufacturers across 40+ countries. This newsletter chases one thing: the gap between the headline and the data. Someone’s always selling hopium. I find it.

The US wants to reshore manufacturing. China wants to keep the industrial lead it has. Both need workers, and both are running short at the same time, but for opposite reasons. I believe those stories are interconnected and worth investigating.

The American Exit

10,000 boomers turn 65 every day, and the wave crests right now, 2024 to 2027. Manufacturing is among the hardest hit: an 11.8% retirement rate, and more than a quarter of the workforce is already 55 or older. By 2035 most of them are gone, and the knowledge walks out with them. A guy with 30 years on a 5-axis took it with him. A welder who spent a career reading blueprint tolerances, gone (the American Welding Society put the gap at 400,000 welders by the mid-2020s, and the average US welder is already near retirement age).

3.8 million manufacturing jobs open up between 2024 and 2033, half of them projected to stay empty. 462,000 sit open today, before a single new reshoring plant has turned on the lights.

Gen Z ran the numbers. Factory-floor production jobs pay about $50,090 a year against a $67,920 national average, the hours are long, the work is physical, and the credentials they want are the ones most high schools stopped teaching in the 1990s, back when the work was leaving anyway. It left. Now it’s coming back, and the pipeline that fed it has been gone for 30 years.

The shortage has a fix, and some people are already running it. Julian Fried at Advanced Metal Research in Hawthorne has hosted a free welding class every Sunday since February. 14 in a row, 70+ people welding for the first time, the youngest 10 years old.

Down the road, Lucas Pacheco is training a new generation on industry-grade equipment at Hawthorne High's School of Manufacturing and Engineering.

We need 1,000 more of both.

And the shops actually winning aren’t fishing in some shrinking pool of credentialed machinists. Zane at NOX:

“the best factory technicians we ever hired had zero experience and didnt know manufacturing was a path they could go down”

Plenty of others, Jim at SendCutSend, Jon at Fabworks, Caleb at OSH Cut, hired their best people with zero manufacturing background (baristas, lawyers, software engineers) and trained them into the best on the floor. The path is wide open. The catch is pay and culture. You have to compensate these people properly and give them a team and a mission worth showing up for, which means the business model underneath has to support a real wage. The shops whose model can't are the ones that will keep telling you the talent doesn't exist.

The Chinese Refusal

Same setup, opposite shape. China has 12.22 million college graduates entering the market this year, a record. Official youth unemployment is 19%, and once you count gig work and underemployment it’s closer to 20%. The factories still can’t find workers.

The word for it is 内卷 (nèijuǎn), involution. Everyone works harder, the bar rises, everyone works harder again, and the reward never moves. A generation looked at 996 for $8 to 10 an hour and opted out. They call it 躺平 (tǎng píng), lying flat: do the minimum, skip the house, skip the marriage, skip the kid. The average workweek is still 48.7 hours. Fixed-asset investment fell 3.8% in 2025, the first drop on record.

Foxconn saw it early and swapped assembly workers for robots, not because robots were cheaper but because Chinese millennials wouldn’t take the shifts their parents took. On April 28 2026, the Ministry of State Security put out a video calling the lying-flat mindset a foreign plot, hostile forces funding influencers to demoralize Chinese youth. When a government has to treat “young people don’t want to grind” as espionage, the contract between state and worker has already snapped.

Foxconn's founder, Terry Gou, said the quiet part back in 2012:

"As human beings are also animals, to manage one million animals gives me a headache."

The robots were always the plan.

On my last trip through Shenzhen and Shanghai, every machine shop owner I talked to brought up lying flat without me asking. So it’s real over there. But the good operators have a workaround: they staff with family, cousins, and old friends, carve out a role for each one, and run the place as a multi-generational family business. That’s the slice of China’s labor base that doesn’t lie flat, and it’s the slice the headline unemployment number can’t see.

The Collision

Two countries and two different demographic crunches, but both betting their industrial future on workers who don’t want the jobs. The US is reshoring chip fabs, battery plants, and defense work into a market already 462,000 people short, with 1.9 million unfillable roles coming by 2033. China is chasing AI supremacy and chip self-sufficiency by 2028 on the backs of a cohort posting 卷 memes and studying abroad to avoid coming home to it.

The robots are supposed to fix this. Eventually they might. But the automation that replaces skilled work (5-axis CNC, precision welding, advanced semiconductor fab) doesn’t exist at scale, and humanoids doing it are a 2030+ story even on the optimistic end. The gap between now and then is 5 to 8 years, and both crises peak right inside it: 2024 to 2033 in the US, 2026 to 2030 in China.

I wrote this week that reindustrialization needs the one thing money can’t buy, multi-decade commitment, and capital patient enough to eat low returns to get there. This is the confirmation. China is promising AI supremacy. America is promising full automation and robotics. Both are betting the machines show up before the workers run out, but the machines aren’t close. So the capital has to sit in this for a decade or more, through thin returns, or it does nothing. An existential threat, or the belief in one, is what carried the US through last time. The open question is whether that belief outlasts a single administration.

I’m Watching This

• Aluminum, up about 70%. Caleb at OSH Cut posted his procurement data: 5052 sheet at 0.08”, roughly $2.40/lb in January 2025, $4.10+ today, while the LME moved only 37%. The rest is the US Midwest premium, where the 50% Section 232 tariff embeds as scarcity.

• US vocational enrollment for 2026 (Q3, from NCES). If community-college manufacturing enrollment still isn't recovering after a decade of decline, the 2033 shortfall numbers are conservative, not pessimistic.

• Foxconn's automation announcement. Three years into replacing assembly workers, the next robot-deployment number isn't about the efficiency. It's a read on how thin the Chinese factory labor pool has actually gotten.

One Last Thing

So, to sum up…. The boomers are retiring with the knowledge in their heads. China’s young won’t take the shifts their parents took. The machines that were supposed to cover for both aren’t ready and won’t be until 2030 at best. The US runs short of people before the first reshored fab opens. China runs short of willing ones before it hits its 2028 targets.

I’m still bullish on reindustrialization. But the shops that win are the ones training their own people from scratch and paying them like it matters, not the ones waiting on a robot or a hiring pool that isn’t coming. And the capital behind them has to be the kind that can wait a decade.

Const

I’m Const. 12 years in US manufacturing. Ran a machine shop, built DigiFabster (quick quoting for job shops), worked with hundreds of manufacturers across 40+ countries. This newsletter chases one thing: the gap between the headline and the data. Someone’s always selling hopium. I find it.

The reindustrialization money is real and it's loud. But almost all of it lands in one place. Defense startups pulled $14.6B in the first five months of 2026, already past all of last year. There are two more layers underneath that one, and I think they decide who actually gets built. Worth walking down all three.

The Top Layer Gets The Checks

Primes, neoprimes, and YC startups are raising like the window shuts tomorrow. Anduril took $5B at a $61B valuation in May. Mach Industries raised $300M. Hadrian hit a $1.6B valuation in January with T. Rowe Price, Founders Fund, and a16z in the round, six months after raising $260M. People in the space will tell you Hadrian and a few others basically have unlimited money right now, and they’re not far off.

Then the old guard moves. On June 16 GM and Lockheed signed an MOU to explore building missile and interceptor parts together. No specific parts named yet (it’s an MOU, not a contract). The last time Detroit retooled for weapons was 1942, when GM turned out tanks and aircraft engines and Ford ran B-24s off Willow Run. That’s the reference point they’re reaching for.

GM can machine the parts. The slow part is getting defense-qualified: ITAR, AS9100, DFARS, a first article floated six months while the government signs off. Allison Giddens, who runs a 30-person precision shop, lived exactly that. Her line from last issue stuck with me:

"Everyone preaches 'resilience' from a conference stage, coming out of the mouths of people who have never had to make payroll."

As primes scale, they spin off thousands of operational problems they can’t solve inside their own walls. Solving those is the opening, becoming the supplier they can’t replace. The capital is watching GM, but I’m convinced the real work is one tier below it.

One Tier Down, The Shops That Climbed

A tier under the primes sit the shops that actually cut metal, and a few of them are winning big. SendCutSend bootstrapped to $200M in revenue, then took $110M from Sequoia and the Collison brothers in May at a $1B valuation, and committed $1B over five years to US capacity. OSH Cut and Fabworks climbed the same road: software first, fast quotes, the small customers nobody else wanted (but now everybody wants them ofc). They built the speed themselves over 5 to 10 years, and speed is the product. Caleb Chamberlain of OSH Cut said it best:

“We’ll never beat China on price. We beat them on speed, we tie them on heavy.”

None of them are defense suppliers though, to be clear. They're not ITAR or CMMC compliant, and the closest they get is prototyping for robotics and space startups. Their customer is the commercial buildout, and that's the side I'd rather build on.

The other way up is outside money. Operators with PE or private capital behind them buy good Tier 1 shops and rebuild them into the fastest suppliers they can hand to Boeing and Lockheed. Bootstrap for a decade on the commercial side, or get bought and re-tooled into a prime supplier. For a legacy shop, I think that second road is the best outcome on the table: capital behind you, and a spot in the supply chain the prime can't reorder around.

The Bottom Layer Can’t Get Funded

However, most shops are on neither road, and that’s the part I’m concerned about. Legacy Tier 1, 2, and 3, running on tribal knowledge, paper travelers, and a quote that lives in one guy’s head. They buy AI tools and get nothing usable back, because the data underneath is a mess. Ben Sheats at The Fabricator said it plainly:

“In the absence of your specific data, a language model will pattern-match against everything it’s ever read about manufacturing in general, and hand you averages dressed up as insights.”

You have to make the shop readable first. Most can’t, and no fund is paying them to try.

So their capital is an SBA loan. Zach Glabman calls it synthetic relationship banking. The shop is illegible to a bank, and he nails why:

"the best-run shop in the county and the worst one might get the same terms, because on paper they look nearly the same."

A federal guarantee stands in for the local banker who used to know them both. The newest tool is MARC, a manufacturing loan line from September 2025, built for the trap that kills these shops: a record order that pays NET 120 while the material bill lands next week (we covered that squeeze last issue). I believe the most boring agency in Washington is doing more for reindustrialization right now than DARPA or the Pentagon. And this is the layer everything above depends on, running on the least money of all.

I’m Watching This

• The FATP gap. em mhouris put it flat on June 20: ten years ago you couldn’t get a part made, today there are ten options for CNC, and assembly never got that treatment. Nobody has built the on-demand version of final assembly, test, and pack. When Tim Cook said China had skilled labor, not cheap labor, this is the part he meant. If the middle layer grows somewhere next, my bet is here.

• Midjourney Medical. On June 18 an AI image company announced a full-body ultrasound scanner. The chip is licensed from Butterfly Network and made by TSMC in Taiwan, and no US manufacturer has been named. American cap table, Taiwanese bill of materials, the same story I wrote last time. I want to know where they actually build it.

• June PMI, out July 1. ISM manufacturing printed 54 in May, fifth straight month of growth, with tariff and input-cost noise sitting underneath. I’m watching whether new orders hold above 50 or the streak finally breaks.

One Last Thing

So, to sum up…. The top gets the term sheets. The middle climbs for a decade, or gets bought and re-tooled into the supplier Boeing depends on. The bottom gets an SBA loan and a long rebuild, or nothing at all. The money is loud at the top. By the bottom you can barely hear it. The bottom is where most of American manufacturing still lives.

I’m bullish on the reindustrialization movement though. That money won’t stay at the top. It reaches the layers beneath too, just not evenly and not as a clean VC check. The hungry, proactive shops will catch it one way or another. The ones that sit still are doomed. That’s the gap, and it’s the whole reason I write this.

Const

I’m Const. 12 years in US manufacturing. Ran a machine shop, built DigiFabster (quick quoting for job shops) and worked with hundreds of manufacturers from over 40 countries.

This newsletter is all about one thing: the gap between what’s in the headlines and what the data shows. Someone is always selling Hopium. I find it.

SpaceX went public this week at about $1.77 trillion, the biggest IPO ever, and more than 4,400 employees woke up millionaires, from senior engineers down to cooks and welders. Add the 400-plus founders it’s seeded over the years and you have the best talent factory American hardtech has ever run.

That’s real, and it’s the best news this industry has had in a while. Then you read the footnotes. About 83% of those founders left space for something else. Bezos just raised $12B to build engineering AI on US soil, with no product and no dataset. And the shop that actually machines the part is running Chinese carbon fiber and getting paid in 120 days. American cap table, Chinese bill of materials.

The Flywheel Is Real

The talent factory isn’t a metaphor. Tom Mueller, employee number one, builds orbital tugs at Impulse; Joe Laurienti’s Ursa Major is 3D-printing the solid rocket motors the Pentagon can’t get enough of, $115M booked last year; three more SpaceX vets started Castelion to make hypersonic missiles cheap and in volume. All real companies and real contracts.

Read it straight, though. The traction sits in defense and aerospace, a real industrial base but a narrow one. The money also runs ahead of the metal. Varda and K2 have flown and booked, but Stoke and Relativity haven’t reached orbit, and the valuations assume they will. It’s a great talent factory. Calling it reindustrialization is a stretch, because it skips the part that actually keeps us tied to China.

Who Owns the Parts

None of these rockets fix what we actually buy from China.

Riley Reese, who runs Arris Composites, said it flat this week: “all US drone companies are flying on Chinese carbon fiber (whether they know it or not).” Raw fiber isn’t covered under Blue UAS yet, so it’s cheap and it’s in everything. Go up to rare-earth magnets, the parts inside every motor, every guidance fin, every drone, and China makes more than 90% of them. You don’t out-found that with a Series A. DoD is buying equity in MP Materials and the Pentagon is banning Chinese magnets in weapons by 2027 precisely because this needs sovereign-scale money and guaranteed offtake, not a founder with a multiple. The flywheel builds the missile but it doesn’t mine, refine, or pour what the missile is made of.

I built quoting software for job shops, so I get to watch how this supply chain actually prices work and actually gets paid. When I walked machine shops in Shenzhen last year, the owner wrote the quote himself at 80 people on the floor, ran it through an LLM for the English, and the part was moving while a comparable US shop was still deciding whether to bid. Speed is the product over there.

Back home, Allison Giddens, who runs a 30-person precision shop, wrote the line that made my week:

Everyone preaches “resilience” from a conference stage, “coming out of the mouths of people who have never had to make payroll.”

For her that’s AS9100, CMMC, DFARS without a lawyer, a First Article floated six months until the government signs off, and a prime asking her to hold a quote for three years. That shop is the base the whole defense buildout stands on, and we keep starving it while the billions get raised everywhere else.

Not every shop can fundraise like SendCutSend or Hadrian, and they shouldn’t have to. Capacity isn’t their problem. Friction is. Days leak out of every step: the RFQ, the bidding, procurement, the handoffs, the quoting. More factories help, sure. But the missing piece is the layer that wires the shops we already have into a supply chain that actually moves.

The Big Raise Builds None of It

The loudest check of the week funds a promise. Bezos unveiled Prometheus at a $41B valuation to build an "artificial general engineer" for the physical world, with no product, no customers, and (the founders admit) no training data, because there's no internet of manufacturing data to scrape. Could be huge. It's also $41B for a thesis that has to manufacture its own raw material first.

Worth remembering who's writing the check. Amazon is the same machine that taught Americans to click "cheapest from Shenzhen" and buried plenty of US makers doing it. Now Bezos wants to underwrite American physical AI. I'll watch the receipts.

While he raises, others ship.

FAW-Volkswagen’s Tianjin plant runs 105 AI applications and went up in 23 months. Andre Wegner of Authentise: “having an AI strategy and using AI might not be the same thing” (those numbers route through Chinese state media, so trust the direction, not the decimals). Maxim Lobovsky of Formlabs flagged the ceiling the leapfrog crowd skips, that digital manufacturing will “make an iPhone 1 before a 100-year-old car.” You still can’t get a big steel casting out of a startup.

I’m Watching This

• Reindustrialize 3.0 hits Detroit this week, Machina Labs' Edward Mehr (ex-SpaceX) on stage. Best read all year on whether the slogans have metal behind them.

• Interest on the federal debt just hit $1.3T over twelve months, about to pass Social Security as the biggest line in the budget. Caleb Chamberlain of OSH Cut: “we print money and light it on fire”.

• 15,000+ satellites are up there now and one company owns close to 10,000, launching most weeks. The debris and traffic math is the quiet bill on the whole low-orbit boom.

  

One Last Thing

The same split shows up everywhere once you look for it. We own the equity, the IP, the brand, the model weights. Someone else owns the factory, the parts, the hands, and the keyboards. The code gets written in Bangalore, the t-shirt sewn in Dhaka, the support call answered in Manila, and the cap table stays in Delaware. China is the sharpest version, but swap in Taiwan for the chips or Vietnam for the assembly and the line reads the same: American cap table, somebody else's bill of materials. The world is more interwoven than the headlines let on. And owning the cap table feels like winning, right up to the week you need the thing it can't make.

Const

This newsletter is all about one thing: the gap between what's in the headlines and what the data shows. Someone is always selling Hopium. I find it.

China has a playbook. Most of us know it cold, but once more: pick a category, scale it at home, drive the cost to the floor, then export the collapse. EVs, humanoids, touch glass. The pattern doesn’t care about the category. It just always lived in hardtech. Until now.

This week it showed up in the one place the market never stops watching: AI. And this round is bigger, because AI isn’t a category. It’s the layer under all the other ones.

All Roads Lead to China. AI Inference Edition.

A developer ran the same job through DeepSeek and Claude and posted the bill: $2 against $265. He saw no difference and switched. Maybe he's exaggerating. Maybe the honest gap is 10x, not 100x. Doesn't matter. 10x is still a precedent, and now it's in software, not steel. The floor is dropping out.

And the market felt it. Roughly a trillion dollars came off the chipmakers this week, on nothing but the fear that demand for expensive compute won’t hold. The whole trade is priced as if AI stays scarce and premium forever. Cheap inference is the crack under that. You don’t deflate a supercycle with a recession. You deflate it when nobody needs the expensive rack anymore.

Run it forward. Cheap inference doesn’t just lower a bill. It decides who owns the layer that sits inside every other piece of software. Whoever’s cheapest becomes the default, and the default isn’t a vendor you swap out. It’s a dependency.

And this dependency is worse than cheap goods ever were. Cheap steel sat in your building and did nothing. Cheap AI reads your files. Every quote and drawing you push through it trains a model you’re renting from the people you’re trying to beat. You cut cost today and feed your competitor tomorrow.

And it's happening right in front of us.

Sources: DeepSeek vs Claude cost test (@TimJayas), chip selloff erases $1T (Seeking Alpha), DeepSeek breaking the inference price floor (CTOL)

The Quoting Gap Is Real. The AI Fix Is Mostly Hopium.

Something caught my eye this week while I was scrolling LinkedIn for the rare non-slop post. Adar Hay, the CEO of Jiga, made a clean point. Send an Asian integrated factory a small, half-baked, early inquiry and they get back to you fast, because they have people who just do quoting. Send the same thing to a US job shop and the owner answers it somewhere between programming a machine and chasing down a payment. The median US shop is 15 people and about $1.5-2M in revenue, across roughly 13,000 of them (the "official" count is higher, I don't buy it, I trust my gut). A shop that size can't keep a quoting team, so it protects the owner's time and only quotes what it's pretty sure it'll win. Rational answer to a broken setup.

To be fair, that’s not the whole story. When I was in China last year, none of the shops looked like the “factory” in that screenshot. Owners were writing their own quotes at 50-100 workers… fast, with real DFM reasoning for why a part gets made one way and not another. English isn’t their first language, so the final proposal goes through an LLM before it lands with the overseas customer. They’ve got more hands and more room to be picky, so they lean on software less. The owner still does the work himself. That’s the nature of this business. The owner is always in the quoting, the invoicing, neck-deep in operations. So the idea that US owners are maxed out and Chinese ones aren’t doesn’t really hold. Work more? Maybe. Want it more? Probably. More competition breathing down their neck? Absolutely.

But to me, here’s where it falls apart. The pitch is that you buy the AI agent and it magically learns how your shop runs. Reads years of messy data and turns it into something useful. Somehow knows the stuff you never wrote down, the tribal knowledge sitting in the head of the operator next to the machine… No, It doesn’t.

A model can’t quote what was never captured, and in a 15-person shop, almost none of it was.

So the software ships and dies on the floor. Not because it’s bad. Because nobody did the change management. The owner who already knows everything (and plenty of them really believe there’s nothing left to learn… that’s the trap) signs the contract and skips the part where you actually talk to the front office and the people running the machines. Miss either one and you get a tool nobody opens and goals that never land, because the floor sees nothing in it for them and quietly works around it.

Sources: Census NAICS 332710, SICCODE 332710, BLS OES 332710

Metal Distribution. Digital or Get Eaten.

Lots of buzz about these guys on X. Mostly dinosaurs. But they're still the ones who set the price your parts sell at.

Same fight, one rung up the supply chain. Your shop decides whether it owns its software or rents it. The distributor selling you the steel is making the exact same call. Interestingly enough, looks like most of them are blowing it.

Everybody’s “building ecommerce.” It didn’t happen today. Many of them had digital experience departments for many years. But most just put up a website and call it an online, digital-first experience. Reality: it's a shopping cart bolted onto a catalog.

However, there's a younger, hungrier one. NOX Metals (YC/Detroit) is building a metal supplier with no humans in the middle that cares and delivers. FAST. Order goes from the website straight to the floor to the truck. No rep, no phone tag, no “let me check with the back and get back to you.” The factory is the product.

Now the other side. Reliance, the biggest service center in the country at about $13.8B in revenue, launched FastMetals in 2020 and ran it through its Yarde Metals arm. It never grew into its own thing and has now been folded back into Yarde (As of May 2026). Reliance never even broke it out in the financials. That tells you how much it mattered to them.

Ryerson went the other way and merged with Olympic Steel (all-stock, closed in February) to become the second-largest service center in the country. Look at what they actually bought though. Scale. Tonnage. Not software. The one incumbent that built something real is Online Metals (thyssenkrupp), a working B2B marketplace, 50+ outside sellers, cut-to-size, shipped to your door. One. Out of the whole old guard.

Here’s the number that says it all. Ecommerce is barely 13% of industrial distribution revenue. The other ~87% still runs on reps, EDI, phone calls, and email threads that look like it’s 2014. The VC money is betting on the digital natives. The incumbent money is betting on getting bigger. Only one of those compounds. So the question was never whether metal goes software-driven. It’s whether a US company builds that software, or just ends up as freight moving through someone else’s. Same answer as every other layer this week. Own it, or become someone else’s traffic.

Sources: Reliance FY2024 results, FastMetals launch, Yarde/FastMetals, Ryerson–Olympic Steel merger close, Online Metals B2B marketplace, Industrial distribution ecommerce ~13%

Own the Means of Production

Let me wrap with something from a guy whose dead hang is 5 minutes 30 seconds. Palantir's Alex Karp.

Cheap compute, cheap quoting software, cheap distribution platforms, all pointing the same way. The winner was never going to be whoever has the cheapest tokens. Cheap is table stakes now. The winner is whoever owns the layer the cheap stuff runs on.

Palantir’s Karp said it without the spin: don’t rent your capability from a lab that “doesn’t care about you at all,” own the means of production. Obviously, he’s selling something when he says it. But he’s also right. A model is raw material, like bar stock. You don’t hand a competitor your bar stock and your tolerances and ask them to run the job for you.

And here’s the thing: you already own the means of production. The machines, the floor, the people, the quoting/ops logic in your head. That’s the reason a shop in Shenzhen can’t copy you overnight. The stakes just doubled. You used to win by owning the factory. Now you win by owning the factory and the AI that runs it.

So treat AI like infrastructure, not magic software. Capture what’s in your people’s heads before it walks out the door. Keep your data on something you control. Buy cheap where it’s throwaway, own it where it’s you. Skip that and you’re not cutting costs. You’re building someone else’s advantage.

Const